RUNYARD.DEV / COMPARE

Compare Devices

Pick two devices · Device B can be wrapped with TurboQuant

Device ANormal
VRAM
RAM
Cores
Backend
Bandwidth~504 GB/s
Device B✦ TurboQuant
Bit-width4× · ~99.5%
safe default · validated on RTX 3060
VRAM
RAM
Cores
Backend
Bandwidth~504 GB/s
Sort
A: 26Tie: 354B: 174
Model
RTX 4070 · Normal
RTX 4070 · ✦ TQ
Qwen3-8B8BUNLOCKED
Reasoning
RUNS GREAT72/100~27.5 tok/s
ctx7K
RUNS GREAT72/100~27.5 tok/s
7K28K
DeepSeek-R1-0528-Qwen3-8B8.2BUNLOCKED
Reasoning
RUNS GREAT72/100~26.8 tok/s
ctx6K
RUNS GREAT73/100~26.8 tok/s
6K26K
granite-3.1-8b-instruct8BUNLOCKED
General
RUNS GREAT69/100~27.5 tok/s
ctx7K
RUNS GREAT69/100~27.5 tok/s
7K28K
Qwen-7B7.7BUNLOCKED
General
RUNS GREAT69/100~28.6 tok/s
ctx8K
RUNS GREAT69/100~28.6 tok/s
8K32K
Qwen1.5-7B7.7BUNLOCKED
General
RUNS GREAT69/100~28.6 tok/s
ctx8K
RUNS GREAT69/100~28.6 tok/s
8K32K
EXAONE-Deep-7.8B7.8BUNLOCKED
General
RUNS GREAT69/100~28.2 tok/s
ctx8K
RUNS GREAT69/100~28.2 tok/s
8K30K
MiMo-7B-Base7.8BUNLOCKED
General
RUNS GREAT69/100~28.2 tok/s
ctx8K
RUNS GREAT69/100~28.2 tok/s
8K30K
saiga_llama3_8b8BUNLOCKED
General
RUNS GREAT69/100~27.5 tok/s
ctx7K
RUNS GREAT69/100~27.5 tok/s
7K8K
Hermes-3-Llama-3.1-8B8BUNLOCKED
General
RUNS GREAT69/100~27.5 tok/s
ctx7K
RUNS GREAT69/100~27.5 tok/s
7K28K
Llama-3.1-Nemotron-Nano-8B-v18BUNLOCKED
General
RUNS GREAT69/100~27.5 tok/s
ctx7K
RUNS GREAT69/100~27.5 tok/s
7K28K
Meta-Llama-3.1-8B-FP88BUNLOCKED
General
RUNS GREAT69/100~27.5 tok/s
ctx7K
RUNS GREAT69/100~27.5 tok/s
7K28K
llava-onevision-qwen2-7b-ov8BUNLOCKED
General
RUNS GREAT69/100~27.5 tok/s
ctx7K
RUNS GREAT69/100~27.5 tok/s
7K28K
granite-4.1-8b8BUNLOCKED
General
RUNS GREAT69/100~27.5 tok/s
ctx7K
RUNS GREAT69/100~27.5 tok/s
7K28K
Qwen3-VL-8B-Thinking8BUNLOCKED
General
RUNS GREAT69/100~27.5 tok/s
ctx7K
RUNS GREAT69/100~27.5 tok/s
7K28K
Qwen3-VL-8B-Instruct8BUNLOCKED
General
RUNS GREAT69/100~27.5 tok/s
ctx7K
RUNS GREAT69/100~27.5 tok/s
7K28K
granite-4.2-8b8BUNLOCKED
General
RUNS GREAT69/100~27.5 tok/s
ctx7K
RUNS GREAT69/100~27.5 tok/s
7K28K
gemma-2-9b-it9BUNLOCKED
General
RUNS GREAT68/100~24.4 tok/s
ctx4K
RUNS GREAT69/100~24.4 tok/s
4K8K
Qwen3-14B-NVFP48.2BUNLOCKED
General
RUNS GREAT68/100~26.8 tok/s
ctx6K
RUNS GREAT69/100~26.8 tok/s
6K26K
Qwen3-8B-Base8.2BUNLOCKED
General
RUNS GREAT68/100~26.8 tok/s
ctx6K
RUNS GREAT69/100~26.8 tok/s
6K26K
Qwen3-8B-AWQ8.2BUNLOCKED
General
RUNS GREAT68/100~26.8 tok/s
ctx6K
RUNS GREAT69/100~26.8 tok/s
6K26K
Qwen3-8B-FP88.2BUNLOCKED
General
RUNS GREAT68/100~26.8 tok/s
ctx6K
RUNS GREAT69/100~26.8 tok/s
6K26K
Qwen3-8B.w8a88.2BUNLOCKED
General
RUNS GREAT68/100~26.8 tok/s
ctx6K
RUNS GREAT69/100~26.8 tok/s
6K26K
Qwen3-8B-FP8-dynamic8.2BUNLOCKED
General
RUNS GREAT68/100~26.8 tok/s
ctx6K
RUNS GREAT69/100~26.8 tok/s
6K26K
LFM2-8B-A1B8.3BMoEUNLOCKED
General
RUNS GREAT68/100~21.2 tok/s
ctx6K
RUNS GREAT69/100~21.2 tok/s
6K25K
NVIDIA-Nemotron-Nano-9B-v28.9BUNLOCKED
General
RUNS GREAT68/100~24.7 tok/s
ctx5K
RUNS GREAT69/100~24.7 tok/s
5K18K
NVIDIA-Nemotron-Nano-9B-v2-Japanese8.9BUNLOCKED
General
RUNS GREAT68/100~24.7 tok/s
ctx5K
RUNS GREAT69/100~24.7 tok/s
5K18K
NVIDIA-Nemotron-Nano-9B-v2-Base8.9BUNLOCKED
General
RUNS GREAT68/100~24.7 tok/s
ctx5K
RUNS GREAT69/100~24.7 tok/s
5K18K
NVIDIA-Nemotron-Nano-9B-v2-FP88.9BUNLOCKED
General
RUNS GREAT68/100~24.7 tok/s
ctx5K
RUNS GREAT69/100~24.7 tok/s
5K18K
QwQ-32B-MLX-8bit9.2BUNLOCKED
General
RUNS GREAT68/100~23.9 tok/s
ctx4K
RUNS GREAT69/100~23.9 tok/s
4K15K
glm-4-9b9.4BUNLOCKED
General
RUNS GREAT68/100~23.4 tok/s
ctx3K
RUNS GREAT69/100~23.4 tok/s
3K8K
Qwen2.5-Coder-32B-Instruct-MLX-8bit9.2BUNLOCKED
Coding
RUNS GREAT67/100~23.9 tok/s
ctx4K
RUNS GREAT68/100~23.9 tok/s
4K15K
Qwen3-Coder-30B-A3B-Instruct-gptq-8bit9.3BMoEUNLOCKED
Coding
RUNS GREAT66/100~18.9 tok/s
ctx4K
RUNS GREAT67/100~18.9 tok/s
4K15K
Llama-3.1-8B-Instruct8BUNLOCKED
Chat
RUNS GREAT63/100~27.5 tok/s
ctx7K
RUNS GREAT63/100~27.5 tok/s
7K28K
Qwen-7B-Chat7.7BUNLOCKED
Chat
RUNS GREAT63/100~28.6 tok/s
ctx8K
RUNS GREAT63/100~28.6 tok/s
8K32K
salamandra-7b-instruct7.8BUNLOCKED
Chat
RUNS GREAT63/100~28.2 tok/s
ctx8K
RUNS GREAT63/100~28.2 tok/s
8K8K
Meta-Llama-3.1-8B-Instruct8BUNLOCKED
Chat
RUNS GREAT63/100~27.5 tok/s
ctx7K
RUNS GREAT63/100~27.5 tok/s
7K28K
Llama-3.1-8B-Instruct-FP88BUNLOCKED
Chat
RUNS GREAT63/100~27.5 tok/s
ctx7K
RUNS GREAT63/100~27.5 tok/s
7K28K
Llama-3-Patronus-Lynx-8B-Instruct-v1.18BUNLOCKED
Chat
RUNS GREAT63/100~27.5 tok/s
ctx7K
RUNS GREAT63/100~27.5 tok/s
7K28K
Meta-Llama-3.1-8B-Instruct-FP88BUNLOCKED
Chat
RUNS GREAT63/100~27.5 tok/s
ctx7K
RUNS GREAT63/100~27.5 tok/s
7K28K
Meta-Llama-3.1-8B-Instruct-quantized.w4a168BUNLOCKED
Chat
RUNS GREAT63/100~27.5 tok/s
ctx7K
RUNS GREAT63/100~27.5 tok/s
7K28K
Meta-Llama-3.1-8B-Instruct-FP8-dynamic8BUNLOCKED
Chat
RUNS GREAT63/100~27.5 tok/s
ctx7K
RUNS GREAT63/100~27.5 tok/s
7K28K
granite-3.3-8b-instruct8.2BUNLOCKED
Chat
RUNS GREAT62/100~26.8 tok/s
ctx6K
RUNS GREAT63/100~26.8 tok/s
6K26K
SDAR-8B-Chat-b328.2BUNLOCKED
Chat
RUNS GREAT62/100~26.8 tok/s
ctx6K
RUNS GREAT63/100~26.8 tok/s
6K26K
Qwen2.5-VL-7B-Instruct8.3BUNLOCKED
Chat
RUNS GREAT62/100~26.5 tok/s
ctx6K
RUNS GREAT63/100~26.5 tok/s
6K25K
rnj-1-instruct8.3BUNLOCKED
Chat
RUNS GREAT62/100~26.5 tok/s
ctx6K
RUNS GREAT63/100~26.5 tok/s
6K25K
Mistral-NeMo-Minitron-8B-Instruct8.4BUNLOCKED
Chat
RUNS GREAT62/100~26.2 tok/s
ctx6K
RUNS GREAT63/100~26.2 tok/s
6K8K
Qwen3-30B-A3B-Instruct-2507-AWQ-8bit9BMoEUNLOCKED
Chat
RUNS GREAT61/100~19.6 tok/s
ctx4K
RUNS GREAT62/100~19.6 tok/s
4K17K
glm-4-9b-chat-hf9.4BUNLOCKED
Chat
RUNS GREAT61/100~23.4 tok/s
ctx3K
RUNS GREAT63/100~23.4 tok/s
3K14K
glm-4-9b-chat9.4BUNLOCKED
Chat
RUNS GREAT61/100~23.4 tok/s
ctx3K
RUNS GREAT63/100~23.4 tok/s
3K14K
Qwen3.5-9B9.7BUNLOCKED
General
RUNS WELL49/100~22.7 tok/s
ctx3K
RUNS WELL51/100~22.7 tok/s
3K11K
Qwen3.5-9B-Base9.7BUNLOCKED
General
RUNS WELL49/100~22.7 tok/s
ctx3K
RUNS WELL51/100~22.7 tok/s
3K11K
Qwen2.5-VL-7B7B
Multimodal
RUNS GREAT73/100~31.4 tok/s
ctx10K
RUNS GREAT73/100~31.4 tok/s
10K32K
DeepSeek-R1-7B7B
Reasoning
RUNS GREAT72/100~31.4 tok/s
ctx10K
RUNS GREAT72/100~31.4 tok/s
10K42K
DeepSeek-R1-Distill-Qwen-7B7.6B
Reasoning
RUNS GREAT72/100~28.9 tok/s
ctx8K
RUNS GREAT72/100~28.9 tok/s
8K33K
Orca-2-7B7B
Reasoning
RUNS GREAT71/100~31.4 tok/s
ctx4K
RUNS GREAT71/100~31.4 tok/s
ctx4K
Qwen2.5-7B-Instruct7B
General
RUNS GREAT69/100~31.4 tok/s
ctx10K
RUNS GREAT69/100~31.4 tok/s
10K42K
Qwen3-32B-quantized.w4a165.7B
General
RUNS GREAT69/100~38.6 tok/s
ctx17K
RUNS GREAT69/100~38.6 tok/s
17K41K
zephyr-7b-beta7.2B
General
RUNS GREAT69/100~30.6 tok/s
ctx10K
RUNS GREAT69/100~30.6 tok/s
10K33K
Mistral-7B-v0.17.2B
General
RUNS GREAT69/100~30.6 tok/s
ctx10K
RUNS GREAT69/100~30.6 tok/s
10K33K
prometheus-7b-v2.07.2B
General
RUNS GREAT69/100~30.6 tok/s
ctx10K
RUNS GREAT69/100~30.6 tok/s
10K33K
xLAM-7b-r7.2B
General
RUNS GREAT69/100~30.6 tok/s
ctx10K
RUNS GREAT69/100~30.6 tok/s
10K33K
dolphin-2.6-mistral-7b7.2B
General
RUNS GREAT69/100~30.6 tok/s
ctx10K
RUNS GREAT69/100~30.6 tok/s
10K33K
Olmo-3-1025-7B7.3B
General
RUNS GREAT69/100~30.1 tok/s
ctx9K
RUNS GREAT69/100~30.1 tok/s
9K37K
Qwen2.5-7B7.6B
General
RUNS GREAT69/100~28.9 tok/s
ctx8K
RUNS GREAT69/100~28.9 tok/s
8K33K
SWE-agent-LM-7B7.6B
General
RUNS GREAT69/100~28.9 tok/s
ctx8K
RUNS GREAT69/100~28.9 tok/s
8K33K
Qwen2-7B7.6B
General
RUNS GREAT69/100~28.9 tok/s
ctx8K
RUNS GREAT69/100~28.9 tok/s
8K33K
VulnLLM-R-7B7.6B
General
RUNS GREAT69/100~28.9 tok/s
ctx8K
RUNS GREAT69/100~28.9 tok/s
8K33K
Hy-MT2-7B7B
General
RUNS GREAT69/100~31.4 tok/s
ctx8K
RUNS GREAT69/100~31.4 tok/s
ctx8K
UI-TARS-1.5-7B7B
General
RUNS GREAT69/100~31.4 tok/s
ctx10K
RUNS GREAT69/100~31.4 tok/s
10K42K
Qwen2.5-Coder-7B7B
Coding
RUNS GREAT68/100~31.4 tok/s
ctx10K
RUNS GREAT68/100~31.4 tok/s
10K42K
granite-3.0-8b-instruct8B
General
RUNS GREAT68/100~27.5 tok/s
ctx4K
RUNS GREAT68/100~27.5 tok/s
ctx4K
CodeLlama-7B-Instruct7B
Coding
RUNS GREAT68/100~31.4 tok/s
ctx10K
RUNS GREAT68/100~31.4 tok/s
10K16K
StarCoder2-7B7B
Coding
RUNS GREAT68/100~31.4 tok/s
ctx10K
RUNS GREAT68/100~31.4 tok/s
10K16K
GLM-4.7-Flash-AWQ-4bit6.4BMoE
General
RUNS GREAT68/100~27.5 tok/s
ctx13K
RUNS GREAT68/100~27.5 tok/s
13K52K
Llammas-base-p1-GPT-4o-human-error-mix-paragraph-GEC6.7B
General
RUNS GREAT68/100~32.8 tok/s
ctx4K
RUNS GREAT67/100~32.8 tok/s
ctx4K
Nous-Hermes-llama-2-7b6.7B
General
RUNS GREAT68/100~32.8 tok/s
ctx4K
RUNS GREAT67/100~32.8 tok/s
ctx4K
Llama-2-7b-hf6.7B
General
RUNS GREAT68/100~32.8 tok/s
ctx4K
RUNS GREAT67/100~32.8 tok/s
ctx4K
CodeLlama-7b-hf6.7B
Coding
RUNS GREAT68/100~32.8 tok/s
ctx12K
RUNS GREAT67/100~32.8 tok/s
12K16K
deepseek-coder-6.7b-instruct6.7B
Coding
RUNS GREAT68/100~32.8 tok/s
ctx12K
RUNS GREAT67/100~32.8 tok/s
12K16K
deepseek-coder-6.7b-base6.7B
Coding
RUNS GREAT68/100~32.8 tok/s
ctx12K
RUNS GREAT67/100~32.8 tok/s
12K16K
Tarsier-7b7.1B
General
RUNS GREAT68/100~31 tok/s
ctx4K
RUNS GREAT68/100~31 tok/s
ctx4K
starcoder2-7b7.2B
Coding
RUNS GREAT68/100~30.6 tok/s
ctx10K
RUNS GREAT68/100~30.6 tok/s
10K16K
falcon-7b7.2B
General
RUNS GREAT68/100~30.6 tok/s
ctx4K
RUNS GREAT67/100~30.6 tok/s
ctx4K
starcoder2-7b-GPTQ7.4B
Coding
RUNS GREAT68/100~29.7 tok/s
ctx9K
RUNS GREAT68/100~29.7 tok/s
9K16K
Qwen2.5-Math-7B7.6B
General
RUNS GREAT68/100~28.9 tok/s
ctx4K
RUNS GREAT68/100~28.9 tok/s
ctx4K
Llama-3.1-8B8B
General
RUNS GREAT68/100~27.5 tok/s
ctx4K
RUNS GREAT68/100~27.5 tok/s
ctx4K
Meta-Llama-3-8B8B
General
RUNS GREAT68/100~27.5 tok/s
ctx4K
RUNS GREAT68/100~27.5 tok/s
ctx4K
llama-3.1-8b-bias-reduced8B
General
RUNS GREAT68/100~27.5 tok/s
ctx4K
RUNS GREAT68/100~27.5 tok/s
ctx4K
Qwen3-235B-A22B235BMoE
Reasoning
RUNS WELL67/100~0.6 tok/s
ctx927
RUNS WELL68/100~0.6 tok/s
9274K
CodeLlama-7b-Instruct-hf6.7B
Coding
RUNS GREAT67/100~32.8 tok/s
ctx4K
RUNS GREAT66/100~32.8 tok/s
ctx4K
OLMoE-1B-7B-01256.9BMoE
General
RUNS GREAT67/100~25.5 tok/s
ctx4K
RUNS GREAT67/100~25.5 tok/s
ctx4K
Qwen2.5-Coder-7B-Instruct7.6B
Coding
RUNS GREAT67/100~28.9 tok/s
ctx8K
RUNS GREAT68/100~28.9 tok/s
8K33K
Qwen2.5-Coder-7B-Instruct-GPTQ-Int47.6B
Coding
RUNS GREAT67/100~28.9 tok/s
ctx8K
RUNS GREAT68/100~28.9 tok/s
8K33K
Qwen2.5-Coder-7B-Instruct-AWQ7.6B
Coding
RUNS GREAT67/100~28.9 tok/s
ctx8K
RUNS GREAT68/100~28.9 tok/s
8K33K
hf-moshiko7.8B
General
RUNS GREAT67/100~28.2 tok/s
ctx3K
RUNS GREAT67/100~28.2 tok/s
ctx3K
Qwen3-30B-A3B30BMoE
Reasoning
RUNS WELL66/100~5.9 tok/s
ctx47K
RUNS WELL66/100~5.9 tok/s
47K128K
OLMo-7B-Instruct7B
General
RUNS GREAT66/100~31.4 tok/s
ctx2K
RUNS GREAT67/100~31.4 tok/s
ctx2K
Falcon-7B-Instruct7B
General
RUNS GREAT66/100~31.4 tok/s
ctx2K
RUNS GREAT67/100~31.4 tok/s
ctx2K
Amber6.7B
General
RUNS GREAT66/100~32.8 tok/s
ctx2K
RUNS GREAT66/100~32.8 tok/s
ctx2K
llama-7b6.7B
General
RUNS GREAT66/100~32.8 tok/s
ctx2K
RUNS GREAT66/100~32.8 tok/s
ctx2K
pythia-6.9b7B
General
RUNS GREAT66/100~31.4 tok/s
ctx2K
RUNS GREAT67/100~31.4 tok/s
ctx2K
Llama-3.2-1B-Instruct1B
Chat
RUNS WELL65/100~220 tok/s
ctx128K
RUNS WELL65/100~220 tok/s
ctx128K
gemma-3-1b-it1B
Chat
RUNS WELL65/100~220 tok/s
ctx33K
RUNS WELL65/100~220 tok/s
ctx33K
Qwen3-4B-Instruct-2507-MLX-8bit1.1B
Chat
RUNS WELL65/100~200 tok/s
ctx158K
RUNS WELL65/100~200 tok/s
158K262K
Qwen3-4B-Instruct-2507-MLX-5bit0.8B
Chat
RUNS WELL64/100~275 tok/s
ctx223K
RUNS WELL64/100~275 tok/s
223K262K
Qwen3-4B-Instruct-2507-MLX-6bit0.9B
Chat
RUNS WELL64/100~244.4 tok/s
ctx196K
RUNS WELL65/100~244.4 tok/s
196K262K
Mistral-7B-Instruct7B
Chat
RUNS GREAT63/100~31.4 tok/s
ctx10K
RUNS GREAT64/100~31.4 tok/s
10K32K
Qwen2.5-0.5B-Instruct0.5B
Chat
RUNS WELL63/100~440 tok/s
ctx128K
RUNS WELL63/100~440 tok/s
ctx128K
Qwen3-4B-Instruct-2507-MLX-4bit0.6B
Chat
RUNS WELL63/100~366.7 tok/s
ctx262K
RUNS WELL63/100~366.7 tok/s
ctx262K
TinyLlama-1.1B-Chat-v1.01.1B
Chat
RUNS WELL63/100~200 tok/s
ctx2K
RUNS WELL63/100~200 tok/s
ctx2K
LFM2.5-1.2B-Instruct1.2B
Chat
RUNS WELL63/100~183.3 tok/s
ctx128K
RUNS WELL63/100~183.3 tok/s
ctx128K
Vikhr-Llama-3.2-1B-Instruct1.2B
Chat
RUNS WELL63/100~183.3 tok/s
ctx131K
RUNS WELL63/100~183.3 tok/s
ctx131K
Mistral-7B-Instruct-v0.27.2B
Chat
RUNS GREAT63/100~30.6 tok/s
ctx10K
RUNS GREAT63/100~30.6 tok/s
10K33K
Mistral-7B-Instruct-v0.37.2B
Chat
RUNS GREAT63/100~30.6 tok/s
ctx10K
RUNS GREAT63/100~30.6 tok/s
10K33K
Mistral-7B-Instruct-v0.3-GPTQ7.2B
Chat
RUNS GREAT63/100~30.6 tok/s
ctx10K
RUNS GREAT63/100~30.6 tok/s
10K33K
Olmo-3-7B-Instruct-SFT7.3B
Chat
RUNS GREAT63/100~30.1 tok/s
ctx9K
RUNS GREAT63/100~30.1 tok/s
9K37K
Falcon3-7B-Instruct7.5B
Chat
RUNS GREAT63/100~29.3 tok/s
ctx9K
RUNS GREAT64/100~29.3 tok/s
9K33K
Qwen2-7B-Instruct7.6B
Chat
RUNS GREAT63/100~28.9 tok/s
ctx8K
RUNS GREAT63/100~28.9 tok/s
8K33K
XCurOS-0.1-8B-Instruct7.6B
Chat
RUNS GREAT63/100~28.9 tok/s
ctx8K
RUNS GREAT63/100~28.9 tok/s
8K33K
Dream-v0-Instruct-7B7.6B
Chat
RUNS GREAT63/100~28.9 tok/s
ctx8K
RUNS GREAT63/100~28.9 tok/s
8K33K
Qwen2.5-7B-Instruct-GPTQ-Int47.6B
Chat
RUNS GREAT63/100~28.9 tok/s
ctx8K
RUNS GREAT63/100~28.9 tok/s
8K33K
Qwen2.5-7B-Instruct-1M7.6B
Chat
RUNS GREAT63/100~28.9 tok/s
ctx8K
RUNS GREAT63/100~28.9 tok/s
8K33K
Yi-6B-Chat6.1B
Chat
RUNS GREAT62/100~36.1 tok/s
ctx4K
RUNS GREAT62/100~36.1 tok/s
ctx4K
Llama-2-7b-chat-hf6.7B
Chat
RUNS GREAT62/100~32.8 tok/s
ctx4K
RUNS GREAT62/100~32.8 tok/s
ctx4K
granite-4.0-h-tiny6.9BMoE
Chat
RUNS GREAT62/100~25.5 tok/s
ctx11K
RUNS GREAT62/100~25.5 tok/s
11K43K
falcon-7b-instruct7.2B
Chat
RUNS GREAT62/100~30.6 tok/s
ctx4K
RUNS GREAT62/100~30.6 tok/s
ctx4K
falcon-mamba-7b-instruct7.3B
Chat
RUNS GREAT62/100~30.1 tok/s
ctx4K
RUNS GREAT62/100~30.1 tok/s
ctx4K
Qwen2.5-Math-7B-Instruct7.6B
Chat
RUNS GREAT62/100~28.9 tok/s
ctx4K
RUNS GREAT62/100~28.9 tok/s
ctx4K
Meta-Llama-3-8B-Instruct8B
Chat
RUNS GREAT62/100~27.5 tok/s
ctx4K
RUNS GREAT62/100~27.5 tok/s
ctx4K
Zamba2-1.2B-instruct1.2B
Chat
RUNS WELL61/100~183.3 tok/s
ctx4K
RUNS WELL61/100~183.3 tok/s
ctx4K
Abliterated-Llama-3.2-1B-Instruct1.2B
Chat
RUNS WELL61/100~183.3 tok/s
ctx4K
RUNS WELL61/100~183.3 tok/s
ctx4K
OLMoE-1B-7B-0125-Instruct6.9BMoE
Chat
RUNS GREAT61/100~25.5 tok/s
ctx4K
RUNS GREAT61/100~25.5 tok/s
ctx4K
Phi-mini-MoE-instruct7.6BMoE
Chat
RUNS GREAT61/100~23.2 tok/s
ctx4K
RUNS GREAT61/100~23.2 tok/s
ctx4K
Qwen3-4B-Thinking-2507-MLX-8bit1.1B
General
RUNS WELL60/100~200 tok/s
ctx158K
RUNS WELL60/100~200 tok/s
158K262K
Qwen3-4B-Thinking-2507-MLX-6bit0.9B
General
RUNS WELL59/100~244.4 tok/s
ctx196K
RUNS WELL60/100~244.4 tok/s
196K262K
LFM2-1.2B1.2B
General
RUNS WELL59/100~183.3 tok/s
ctx128K
RUNS WELL59/100~183.3 tok/s
ctx128K
LFM2.5-1.2B-Thinking1.2B
General
RUNS WELL59/100~183.3 tok/s
ctx128K
RUNS WELL59/100~183.3 tok/s
ctx128K
LFM2.5-1.2B-JP1.2B
General
RUNS WELL59/100~183.3 tok/s
ctx128K
RUNS WELL59/100~183.3 tok/s
ctx128K
LFM2-1.2B-Extract1.2B
General
RUNS WELL59/100~183.3 tok/s
ctx128K
RUNS WELL59/100~183.3 tok/s
ctx128K
LFM2-1.2B-MLX-bf161.2B
General
RUNS WELL59/100~183.3 tok/s
ctx128K
RUNS WELL59/100~183.3 tok/s
ctx128K
Ilama-3.2-1B1.2B
General
RUNS WELL59/100~183.3 tok/s
ctx131K
RUNS WELL59/100~183.3 tok/s
ctx131K
CyberXP_Agent_Llama_3.2_1B1.2B
General
RUNS WELL59/100~183.3 tok/s
ctx131K
RUNS WELL59/100~183.3 tok/s
ctx131K
Orpo-Llama-3.2-1B-15k1.2B
General
RUNS WELL59/100~183.3 tok/s
ctx131K
RUNS WELL59/100~183.3 tok/s
ctx131K
Qwen3-4B-MLX-4bit0.6B
General
RUNS WELL58/100~366.7 tok/s
ctx66K
RUNS WELL58/100~366.7 tok/s
ctx66K
Qwen3-4B-Thinking-2507-MLX-4bit0.6B
General
RUNS WELL58/100~366.7 tok/s
ctx262K
RUNS WELL58/100~366.7 tok/s
ctx262K
Qwen3-8B-speculator.eagle31B
General
RUNS WELL58/100~220 tok/s
ctx4K
RUNS WELL59/100~220 tok/s
ctx4K
pythia-1b1.1B
General
RUNS WELL58/100~200 tok/s
ctx2K
RUNS WELL58/100~200 tok/s
ctx2K
Qwen2.5-1.5B-Instruct1.5B
Chat
RUNS WELL57/100~146.7 tok/s
ctx111K
RUNS WELL57/100~146.7 tok/s
111K128K
Qwen3-4B4B
Reasoning
RUNS WELL57/100~55 tok/s
ctx31K
RUNS WELL56/100~55 tok/s
31K124K
Qwen3-4B-DFlash-b160.5B
General
RUNS WELL57/100~440 tok/s
ctx41K
RUNS WELL57/100~440 tok/s
ctx41K
Llama-3.2-1B1.2B
General
RUNS WELL57/100~183.3 tok/s
ctx4K
RUNS WELL57/100~183.3 tok/s
ctx4K
Jan-nano-AWQ1.3B
General
RUNS WELL57/100~169.2 tok/s
ctx41K
RUNS WELL58/100~169.2 tok/s
ctx41K
EXAONE-4.0-1.2B1.3B
General
RUNS WELL57/100~169.2 tok/s
ctx66K
RUNS WELL58/100~169.2 tok/s
ctx66K
Qwen3-8B-MLX-4bit1.3B
General
RUNS WELL57/100~169.2 tok/s
ctx41K
RUNS WELL58/100~169.2 tok/s
ctx41K
plamo-2-1b1.3B
General
RUNS WELL57/100~169.2 tok/s
ctx131K
RUNS WELL58/100~169.2 tok/s
131K523K
Llama-3.2-1B-Instruct-FP81.5B
Chat
RUNS WELL57/100~146.7 tok/s
ctx111K
RUNS WELL57/100~146.7 tok/s
111K131K
Llama-3.2-1B-Instruct-FP8-dynamic1.5B
Chat
RUNS WELL57/100~146.7 tok/s
ctx111K
RUNS WELL57/100~146.7 tok/s
111K131K
Qwen2-1.5B-Instruct1.5B
Chat
RUNS WELL57/100~146.7 tok/s
ctx33K
RUNS WELL57/100~146.7 tok/s
ctx33K
Qwen2-1.5B-Instruct-FP81.5B
Chat
RUNS WELL57/100~146.7 tok/s
ctx33K
RUNS WELL57/100~146.7 tok/s
ctx33K
gpt_bigcode-santacoder1.1B
Coding
RUNS WELL56/100~200 tok/s
ctx2K
RUNS WELL56/100~200 tok/s
ctx2K
OLMo-1B-hf1.2B
General
RUNS WELL56/100~183.3 tok/s
ctx2K
RUNS WELL56/100~183.3 tok/s
ctx2K
llama-3.2-1b-code-instruct1.2B
Coding
RUNS WELL56/100~183.3 tok/s
ctx131K
RUNS WELL57/100~183.3 tok/s
ctx131K
starvector-1b-im2svg1.4B
General
RUNS WELL56/100~157.1 tok/s
ctx8K
RUNS WELL56/100~157.1 tok/s
ctx8K
OLMo-2-0425-1B-Instruct1.5B
Chat
RUNS WELL56/100~146.7 tok/s
ctx4K
RUNS WELL56/100~146.7 tok/s
ctx4K
Qwen2.5-1.5B1.5B
General
RUNS WELL56/100~146.7 tok/s
ctx111K
RUNS WELL55/100~146.7 tok/s
111K131K
Qwen2-1.5B1.5B
General
RUNS WELL56/100~146.7 tok/s
ctx111K
RUNS WELL55/100~146.7 tok/s
111K131K
Qwen2.5-Math-1.5B-Instruct1.5B
Chat
RUNS WELL56/100~146.7 tok/s
ctx4K
RUNS WELL56/100~146.7 tok/s
ctx4K
xLAM-2-1b-fc-r1.5B
General
RUNS WELL56/100~146.7 tok/s
ctx33K
RUNS WELL55/100~146.7 tok/s
ctx33K
qwen-base-invoicev1.01-1.5B1.5B
General
RUNS WELL56/100~146.7 tok/s
ctx33K
RUNS WELL55/100~146.7 tok/s
ctx33K
Llama-4-Maverick400BMoE
General
RUNS WELL55/100~0.4 tok/s
ctx2K
RUNS WELL57/100~0.4 tok/s
2K7K
bge-m30.57B
Embedding
RUNS WELL55/100~386 tok/s
ctx8K
RUNS WELL55/100~386 tok/s
ctx8K
LFM2-VL-1.6B1.6B
General
RUNS WELL55/100~137.5 tok/s
ctx103K
RUNS WELL55/100~137.5 tok/s
103K128K
LFM2.5-VL-1.6B1.6B
General
RUNS WELL55/100~137.5 tok/s
ctx103K
RUNS WELL55/100~137.5 tok/s
103K128K
stablelm-2-1_6b-chat1.6B
Chat
RUNS WELL55/100~137.5 tok/s
ctx4K
RUNS WELL55/100~137.5 tok/s
ctx4K
Qwen3.5-4B4.7B
General
RUNS WELL55/100~46.8 tok/s
ctx24K
RUNS WELL54/100~46.8 tok/s
24K95K
Qwen3.5-4B-Base4.7B
General
RUNS WELL55/100~46.8 tok/s
ctx24K
RUNS WELL54/100~46.8 tok/s
24K95K
Qwen3-8B-NVFP44.7B
General
RUNS WELL55/100~46.8 tok/s
ctx24K
RUNS WELL54/100~46.8 tok/s
24K41K
NVIDIA-Nemotron-3-Nano-30B-A3B-AWQ5.1B
General
RUNS WELL55/100~43.1 tok/s
ctx21K
RUNS WELL55/100~43.1 tok/s
21K83K
Qwen3-32B-MLX-4bit5.1B
General
RUNS WELL55/100~43.1 tok/s
ctx21K
RUNS WELL55/100~43.1 tok/s
21K41K
QwQ-32B-MLX-4bit5.1B
General
RUNS WELL55/100~43.1 tok/s
ctx21K
RUNS WELL55/100~43.1 tok/s
21K83K
DeepSeek-V2-Chat236BMoE
General
RUNS WELL54/100~0.6 tok/s
ctx1K
RUNS WELL56/100~0.6 tok/s
1K5K
DeepSeek-R1-0528-Qwen3-8B-MLX-4bit1.3B
Reasoning
RUNS WELL54/100~169.2 tok/s
ctx131K
RUNS WELL54/100~169.2 tok/s
131K131K
gpt-neo-1.3B1.4B
General
RUNS WELL54/100~157.1 tok/s
ctx2K
RUNS WELL54/100~157.1 tok/s
ctx2K
phi-1_51.4B
General
RUNS WELL54/100~157.1 tok/s
ctx2K
RUNS WELL54/100~157.1 tok/s
ctx2K
LFM2.5-Audio-1.5B1.5B
General
RUNS WELL54/100~146.7 tok/s
ctx4K
RUNS WELL54/100~146.7 tok/s
ctx4K
OLMo-2-0425-1B1.5B
General
RUNS WELL54/100~146.7 tok/s
ctx4K
RUNS WELL54/100~146.7 tok/s
ctx4K
Qwen2.5-Math-1.5B1.5B
General
RUNS WELL54/100~146.7 tok/s
ctx4K
RUNS WELL54/100~146.7 tok/s
ctx4K
SmolLM2-1.7B1.7B
General
RUNS WELL54/100~129.4 tok/s
ctx8K
RUNS WELL54/100~129.4 tok/s
ctx8K
Nanbeige4.1-3B-AWQ-8bit1.7B
General
RUNS WELL54/100~129.4 tok/s
ctx96K
RUNS WELL54/100~129.4 tok/s
96K262K
Qwen3-1.7B-Base1.7B
General
RUNS WELL54/100~129.4 tok/s
ctx33K
RUNS WELL54/100~129.4 tok/s
ctx33K
Qwen3-1.7B-MLX-bf161.7B
General
RUNS WELL54/100~129.4 tok/s
ctx41K
RUNS WELL54/100~129.4 tok/s
ctx41K
Qwen2.5-1.5B-Instruct-AWQ1.8B
Chat
RUNS WELL54/100~122.2 tok/s
ctx33K
RUNS WELL54/100~122.2 tok/s
ctx33K
Qwen2-1.5B-Instruct-AWQ1.8B
Chat
RUNS WELL54/100~122.2 tok/s
ctx33K
RUNS WELL54/100~122.2 tok/s
ctx33K
Qwen2-1.5B-Instruct-GPTQ-Int41.8B
Chat
RUNS WELL54/100~122.2 tok/s
ctx33K
RUNS WELL54/100~122.2 tok/s
ctx33K
Qwen2.5-1.5B-quantized.w8a81.8B
General
RUNS WELL54/100~122.2 tok/s
ctx33K
RUNS WELL54/100~122.2 tok/s
ctx33K
Qwen1.5-1.8B-Chat1.8B
Chat
RUNS WELL54/100~122.2 tok/s
ctx33K
RUNS WELL54/100~122.2 tok/s
ctx33K
Qwen3-14B-MLX-8bit4.2B
General
RUNS WELL54/100~52.4 tok/s
ctx29K
RUNS WELL54/100~52.4 tok/s
29K41K
Qwen3-4B-SafeRL4.4B
General
RUNS WELL54/100~50 tok/s
ctx27K
RUNS WELL54/100~50 tok/s
27K41K
Qwen3-4B-FP84.4B
General
RUNS WELL54/100~50 tok/s
ctx27K
RUNS WELL54/100~50 tok/s
27K41K
Nemotron-H-4B-Base-8K4.5B
General
RUNS WELL54/100~48.9 tok/s
ctx8K
RUNS WELL54/100~48.9 tok/s
ctx8K
Qwen2.5-Coder-32B-Instruct-MLX-4bit5.1B
Coding
RUNS WELL54/100~43.1 tok/s
ctx21K
RUNS WELL54/100~43.1 tok/s
21K33K
Hy-MT2-1.8B1.8B
Chat
RUNS WELL54/100~122.2 tok/s
ctx8K
RUNS WELL54/100~122.2 tok/s
ctx8K
Qwen2.5-Coder-1.5B1.5B
Coding
RUNS WELL53/100~146.7 tok/s
ctx111K
RUNS WELL53/100~146.7 tok/s
111K128K
gemma-2-2b-it2B
Chat
RUNS WELL53/100~110 tok/s
ctx8K
RUNS WELL53/100~110 tok/s
ctx8K
granite-3.1-2b-instruct2B
General
RUNS WELL53/100~110 tok/s
ctx79K
RUNS WELL53/100~110 tok/s
79K128K
nomic-embed-text-v1.50.14B
Embedding
RUNS WELL53/100~1571.4 tok/s
ctx8K
RUNS WELL53/100~1571.4 tok/s
ctx8K
pythia-1.4b1.5B
General
RUNS WELL53/100~146.7 tok/s
ctx2K
RUNS WELL53/100~146.7 tok/s
ctx2K
Qwen2.5-Coder-1.5B-Instruct1.5B
Coding
RUNS WELL53/100~146.7 tok/s
ctx33K
RUNS WELL53/100~146.7 tok/s
ctx33K
Minnow-Math-1.5B1.6B
General
RUNS WELL53/100~137.5 tok/s
ctx4K
RUNS WELL54/100~137.5 tok/s
ctx4K
QVikhr-3-1.7B-Instruction-noreasoning1.7B
Reasoning
RUNS WELL53/100~129.4 tok/s
ctx41K
RUNS WELL53/100~129.4 tok/s
ctx41K
bloom-1b71.7B
General
RUNS WELL53/100~129.4 tok/s
ctx4K
RUNS WELL53/100~129.4 tok/s
ctx4K
DeepSeek-R1-Distill-Qwen-1.5B1.8B
Reasoning
RUNS WELL53/100~122.2 tok/s
ctx90K
RUNS WELL53/100~122.2 tok/s
90K131K
Qwen3-1.7B2B
General
RUNS WELL53/100~110 tok/s
ctx41K
RUNS WELL53/100~110 tok/s
ctx41K
Qwen3-1.7B-FP82B
General
RUNS WELL53/100~110 tok/s
ctx41K
RUNS WELL53/100~110 tok/s
ctx41K
Phi-4-reasoning-plus-MLX-4bit2.3B
Reasoning
RUNS WELL53/100~95.7 tok/s
ctx33K
RUNS WELL53/100~95.7 tok/s
ctx33K
DeepSeek-R1-0528-Qwen3-8B-MLX-8bit2.3B
Reasoning
RUNS WELL53/100~95.7 tok/s
ctx66K
RUNS WELL53/100~95.7 tok/s
66K131K
HTML-Pruner-Phi-3.8B3.8B
General
RUNS WELL53/100~57.9 tok/s
ctx34K
RUNS WELL53/100~57.9 tok/s
34K131K
Nanbeige4.1-3B3.9B
General
RUNS WELL53/100~56.4 tok/s
ctx32K
RUNS WELL53/100~56.4 tok/s
32K129K
Qwen3-4B-Base4B
General
RUNS WELL53/100~55 tok/s
ctx31K
RUNS WELL53/100~55 tok/s
31K33K
Qwen3-4B-Thinking-25074B
General
RUNS WELL53/100~55 tok/s
ctx31K
RUNS WELL53/100~55 tok/s
31K124K
Qwen3-4B-AWQ4B
General
RUNS WELL53/100~55 tok/s
ctx31K
RUNS WELL53/100~55 tok/s
31K41K
Jan-v1-4B4B
General
RUNS WELL53/100~55 tok/s
ctx31K
RUNS WELL53/100~55 tok/s
31K124K
Jan-nano-128k4B
General
RUNS WELL53/100~55 tok/s
ctx31K
RUNS WELL53/100~55 tok/s
31K124K
VLM2Vec-Full4.1B
General
RUNS WELL53/100~53.7 tok/s
ctx30K
RUNS WELL54/100~53.7 tok/s
30K119K
Qwen3-Coder-30B-A3B-Instruct-AWQ-4bit5.3BMoE
Coding
RUNS WELL53/100~33.2 tok/s
ctx19K
RUNS WELL53/100~33.2 tok/s
19K77K
Llama-4-Scout17BMoE
General
RUNS WELL52/100~10.4 tok/s
ctx47K
RUNS WELL52/100~10.4 tok/s
47K188K
Mixtral-8x7B-Instruct46.7BMoE
General
RUNS WELL52/100~3.7 tok/s
ctx2K
RUNS WELL55/100~3.7 tok/s
2K7K
granite-3.0-2b-instruct2B
General
RUNS WELL52/100~110 tok/s
ctx4K
RUNS WELL52/100~110 tok/s
ctx4K
SmolLM-1.7B1.7B
General
RUNS WELL52/100~129.4 tok/s
ctx2K
RUNS WELL52/100~129.4 tok/s
ctx2K
Qwen2.5-Coder-1.5B-Instruct-AWQ1.8B
Coding
RUNS WELL52/100~122.2 tok/s
ctx33K
RUNS WELL52/100~122.2 tok/s
ctx33K
Qwen3.5-2B2.3B
General
RUNS WELL52/100~95.7 tok/s
ctx66K
RUNS WELL52/100~95.7 tok/s
66K262K
Qwen3.5-2B-Base2.3B
General
RUNS WELL52/100~95.7 tok/s
ctx66K
RUNS WELL52/100~95.7 tok/s
66K262K
Qwen3-8B-MLX-8bit2.3B
General
RUNS WELL52/100~95.7 tok/s
ctx41K
RUNS WELL52/100~95.7 tok/s
ctx41K
Qwen3-14B-MLX-4bit2.3B
General
RUNS WELL52/100~95.7 tok/s
ctx41K
RUNS WELL52/100~95.7 tok/s
ctx41K
LFM2-2.6B2.6B
General
RUNS WELL52/100~84.6 tok/s
ctx57K
RUNS WELL52/100~84.6 tok/s
57K128K
LFM2-2.6B-Exp2.6B
General
RUNS WELL52/100~84.6 tok/s
ctx57K
RUNS WELL52/100~84.6 tok/s
57K128K
T-lite-it-1.0_Q4_02.9B
General
RUNS WELL52/100~75.9 tok/s
ctx33K
RUNS WELL52/100~75.9 tok/s
ctx33K
LFM2-VL-3B3B
General
RUNS WELL52/100~73.3 tok/s
ctx47K
RUNS WELL52/100~73.3 tok/s
47K128K
SmolLM3-3B3.1B
General
RUNS WELL52/100~71 tok/s
ctx45K
RUNS WELL52/100~71 tok/s
45K66K
SmolLM3-3B-Base3.1B
General
RUNS WELL52/100~71 tok/s
ctx45K
RUNS WELL52/100~71 tok/s
45K66K
Qwen2.5-3B3.1B
General
RUNS WELL52/100~71 tok/s
ctx33K
RUNS WELL52/100~71 tok/s
ctx33K
xLAM-2-3b-fc-r3.1B
General
RUNS WELL52/100~71 tok/s
ctx33K
RUNS WELL52/100~71 tok/s
ctx33K
Hermes-3-Llama-3.2-3B3.2B
General
RUNS WELL52/100~68.8 tok/s
ctx43K
RUNS WELL52/100~68.8 tok/s
43K131K
Qwen2.5-Coder-14B-Instruct-MLX-8bit4.2B
Coding
RUNS WELL52/100~52.4 tok/s
ctx29K
RUNS WELL52/100~52.4 tok/s
29K33K
xflux_text_encoders4.8B
Coding
RUNS WELL52/100~45.8 tok/s
ctx4K
RUNS WELL52/100~45.8 tok/s
ctx4K
bge-large-en-v1.50.34B
Embedding
RUNS WELL51/100~647.1 tok/s
ctx512
RUNS WELL51/100~647.1 tok/s
ctx512
h2ovl-mississippi-2b2.2B
General
RUNS WELL51/100~100 tok/s
ctx4K
RUNS WELL51/100~100 tok/s
ctx4K
EXAONE-3.5-2.4B-Instruct2.4B
Chat
RUNS WELL51/100~91.7 tok/s
ctx33K
RUNS WELL51/100~91.7 tok/s
ctx33K
gemma-1.1-2b-it2.5B
General
RUNS WELL51/100~88 tok/s
ctx4K
RUNS WELL51/100~88 tok/s
ctx4K
gemma-2-2b-jpn-it2.6B
General
RUNS WELL51/100~84.6 tok/s
ctx4K
RUNS WELL50/100~84.6 tok/s
ctx4K
stablelm-3b-4e1t2.8B
General
RUNS WELL51/100~78.6 tok/s
ctx4K
RUNS WELL51/100~78.6 tok/s
ctx4K
granite-4.0-h-micro3.2BMoE
General
RUNS WELL51/100~55 tok/s
ctx43K
RUNS WELL51/100~55 tok/s
43K131K
Llama-3.2-3B3.2B
General
RUNS WELL51/100~68.8 tok/s
ctx4K
RUNS WELL51/100~68.8 tok/s
ctx4K
PowerLM-3b3.5B
General
RUNS WELL51/100~62.9 tok/s
ctx4K
RUNS WELL51/100~62.9 tok/s
ctx4K
Qwen3-Coder-30B-A3B-Instruct-AWQ4.6BMoE
Coding
RUNS WELL51/100~38.3 tok/s
ctx25K
RUNS WELL51/100~38.3 tok/s
25K99K
Qwen2.5-VL-7B-Instruct-NVFP45B
Chat
RUNS WELL51/100~44 tok/s
ctx21K
RUNS WELL50/100~44 tok/s
21K86K
Phi-4-multimodal-instruct5.6B
Chat
RUNS WELL51/100~39.3 tok/s
ctx17K
RUNS WELL52/100~39.3 tok/s
17K69K
Llama-3.2-3B-Instruct3B
Chat
RUNS WELL50/100~73.3 tok/s
ctx47K
RUNS WELL50/100~73.3 tok/s
47K128K
Qwen2.5-3B-Instruct3B
Chat
RUNS WELL50/100~73.3 tok/s
ctx47K
RUNS WELL50/100~73.3 tok/s
47K128K
gemma-3-4b-it4B
Chat
RUNS WELL50/100~55 tok/s
ctx31K
RUNS WELL50/100~55 tok/s
31K124K
Phi-3.5-mini-instruct3.8B
Chat
RUNS WELL50/100~57.9 tok/s
ctx34K
RUNS WELL50/100~57.9 tok/s
34K128K
Phi-4-mini3.8B
Chat
RUNS WELL50/100~57.9 tok/s
ctx34K
RUNS WELL50/100~57.9 tok/s
34K128K
DeepSeek-Coder-V2-16B16BMoE
Coding
RUNS WELL50/100~11 tok/s
ctx63K
RUNS WELL50/100~11 tok/s
63K128K
StarCoder2-3B3B
Coding
RUNS WELL50/100~73.3 tok/s
ctx16K
RUNS WELL50/100~73.3 tok/s
ctx16K
Qwen2.5-Coder-14B-Instruct-MLX-4bit2.3B
Coding
RUNS WELL50/100~95.7 tok/s
ctx33K
RUNS WELL50/100~95.7 tok/s
ctx33K
gpt-neo-2.7B2.7B
General
RUNS WELL50/100~81.5 tok/s
ctx2K
RUNS WELL50/100~81.5 tok/s
ctx2K
phi-22.8B
General
RUNS WELL50/100~78.6 tok/s
ctx2K
RUNS WELL49/100~78.6 tok/s
ctx2K
pythia-2.8b2.9B
General
RUNS WELL50/100~75.9 tok/s
ctx2K
RUNS WELL50/100~75.9 tok/s
ctx2K
starcoder2-3b3B
Coding
RUNS WELL50/100~73.3 tok/s
ctx16K
RUNS WELL50/100~73.3 tok/s
ctx16K
Qwen2.5-Coder-3B-Instruct3.1B
Coding
RUNS WELL50/100~71 tok/s
ctx33K
RUNS WELL50/100~71 tok/s
ctx33K
Qwen2.5-Coder-3B3.1B
Coding
RUNS WELL50/100~71 tok/s
ctx33K
RUNS WELL50/100~71 tok/s
ctx33K
Qwen2.5-VL-3B-Instruct3.8B
Chat
RUNS WELL50/100~57.9 tok/s
ctx34K
RUNS WELL50/100~57.9 tok/s
34K128K
Qwen3-4B-Instruct-25074B
Chat
RUNS WELL50/100~55 tok/s
ctx31K
RUNS WELL50/100~55 tok/s
31K124K
Qwen3-4B-Instruct-2507-GPTQ-Int44B
Chat
RUNS WELL50/100~55 tok/s
ctx31K
RUNS WELL50/100~55 tok/s
31K124K
Qwen3-4B-Instruct-2507-FP84.4B
Chat
RUNS WELL50/100~50 tok/s
ctx27K
RUNS WELL50/100~50 tok/s
27K107K
Nemotron-H-4B-Instruct-128K4.5B
Chat
RUNS WELL50/100~48.9 tok/s
ctx26K
RUNS WELL50/100~48.9 tok/s
26K103K
Qwen3-30B-A3B-Instruct-2507-AWQ-4bit5.3BMoE
Chat
RUNS WELL50/100~33.2 tok/s
ctx19K
RUNS WELL50/100~33.2 tok/s
19K77K
LFM2.5-2.6B2.6B
Chat
RUNS WELL50/100~84.6 tok/s
ctx57K
RUNS WELL50/100~84.6 tok/s
57K66K
Nemotron-3.5-Content-Safety4B
Chat
RUNS WELL50/100~55 tok/s
ctx31K
RUNS WELL50/100~55 tok/s
31K124K
granite-4.0-h-tiny-AWQ-4bit2BMoE
Chat
RUNS WELL49/100~88 tok/s
ctx79K
RUNS WELL49/100~88 tok/s
79K131K
PowerMoE-3b3.4BMoE
General
RUNS WELL49/100~51.8 tok/s
ctx4K
RUNS WELL50/100~51.8 tok/s
ctx4K
Qwen2.5-3B-Instruct-AWQ3.4B
Chat
RUNS WELL49/100~64.7 tok/s
ctx33K
RUNS WELL49/100~64.7 tok/s
ctx33K
granite-3b-code-base-2k3.5B
Coding
RUNS WELL49/100~62.9 tok/s
ctx2K
RUNS WELL49/100~62.9 tok/s
ctx2K
Llama-3.2-3B-Instruct-FP83.6B
Chat
RUNS WELL49/100~61.1 tok/s
ctx36K
RUNS WELL50/100~61.1 tok/s
36K131K
Qwen3-32B32B
Reasoning
DECENT48/100~2.9 tok/s
ctx
DECENT48/100~2.9 tok/s
ctx
Phi-3-mini-4k3.8B
Chat
RUNS WELL48/100~57.9 tok/s
ctx4K
RUNS WELL48/100~57.9 tok/s
ctx4K
DeepSeek-R1-32B32B
Reasoning
DECENT48/100~2.9 tok/s
ctx
DECENT48/100~2.9 tok/s
ctx
Phi-3-mini-4k-instruct-AWQ3.8B
Chat
RUNS WELL48/100~57.9 tok/s
ctx4K
RUNS WELL48/100~57.9 tok/s
ctx4K
Phi-3-mini-4k-instruct-gptq-4bit3.8B
Chat
RUNS WELL48/100~57.9 tok/s
ctx4K
RUNS WELL48/100~57.9 tok/s
ctx4K
Qwen3-30B-A3B-Instruct-2507-AWQ4.6BMoE
Chat
RUNS WELL48/100~38.3 tok/s
ctx25K
RUNS WELL48/100~38.3 tok/s
25K99K
Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled27.8B
Reasoning
DECENT48/100~3.6 tok/s
ctx
DECENT48/100~3.6 tok/s
ctx
DeepSeek-R1-Distill-Qwen-32B32.8B
Reasoning
DECENT48/100~2.9 tok/s
ctx
DECENT49/100~2.9 tok/s
ctx
OpenReasoning-Nemotron-32B32.8B
Reasoning
DECENT48/100~2.9 tok/s
ctx
DECENT49/100~2.9 tok/s
ctx
Phi-tiny-MoE-instruct3.8BMoE
Chat
RUNS WELL46/100~46.3 tok/s
ctx4K
RUNS WELL46/100~46.3 tok/s
ctx4K
DeepSeek-R1-Distill-Qwen-14B14.8B
Reasoning
DECENT46/100~7.4 tok/s
ctx
DECENT46/100~7.4 tok/s
ctx
Qwen3-14B14B
Reasoning
DECENT45/100~7.9 tok/s
ctx
DECENT45/100~7.9 tok/s
ctx
Phi-414B
Reasoning
DECENT45/100~7.9 tok/s
ctx
DECENT45/100~7.9 tok/s
ctx
DeepSeek-R1-14B14B
Reasoning
DECENT45/100~7.9 tok/s
ctx
DECENT45/100~7.9 tok/s
ctx
Orca-2-13B13B
Reasoning
DECENT45/100~8.5 tok/s
ctx
DECENT45/100~8.5 tok/s
ctx
HyperCLOVAX-SEED-Omni-8B10.7B
General
RUNS WELL43/100~20.6 tok/s
ctx943
RUNS WELL46/100~20.6 tok/s
9434K
Llama-3.2-11B-Vision11B
Multimodal
DECENT40/100~10 tok/s
ctx454
DECENT42/100~10 tok/s
4542K
Llama-3.2-11B-Vision-Instruct10.7B
Chat
RUNS WELL37/100~20.6 tok/s
ctx943
RUNS WELL39/100~20.6 tok/s
9434K
SOLAR-10.7B-Instruct-v1.010.7B
Chat
RUNS WELL37/100~20.6 tok/s
ctx943
RUNS WELL39/100~20.6 tok/s
9434K
CodeLlama-34B-Instruct34B
Coding
DECENT36/100~2.8 tok/s
ctx
DECENT35/100~2.8 tok/s
ctx
CodeLlama-34b-Instruct-hf33.7B
Coding
DECENT36/100~2.8 tok/s
ctx
DECENT35/100~2.8 tok/s
ctx
Qwen2.5-Coder-32B32B
Coding
DECENT35/100~2.9 tok/s
ctx
DECENT35/100~2.9 tok/s
ctx
Qwen3-Coder-30B-A3B-Instruct30.5BMoE
Coding
DECENT35/100~2.5 tok/s
ctx
DECENT35/100~2.5 tok/s
ctx
Qwen3-Coder-30B-A3B-Instruct-MLX-4bit30.5BMoE
Coding
DECENT35/100~2.5 tok/s
ctx
DECENT35/100~2.5 tok/s
ctx
Qwen3-Coder-30B-A3B-Instruct-MLX-5bit30.5BMoE
Coding
DECENT35/100~2.5 tok/s
ctx
DECENT35/100~2.5 tok/s
ctx
Qwen3-Coder-30B-A3B-Instruct-MLX-8bit30.5BMoE
Coding
DECENT35/100~2.5 tok/s
ctx
DECENT35/100~2.5 tok/s
ctx
Qwen3-Coder-30B-A3B-Instruct-MLX-6bit30.5BMoE
Coding
DECENT35/100~2.5 tok/s
ctx
DECENT35/100~2.5 tok/s
ctx
Qwen3-Coder-30B-A3B-Instruct-gptq-4bit30.5BMoE
Coding
DECENT35/100~2.5 tok/s
ctx
DECENT35/100~2.5 tok/s
ctx
Qwen3-Coder-30B-A3B-Instruct-FP830.5BMoE
Coding
DECENT35/100~2.5 tok/s
ctx
DECENT35/100~2.5 tok/s
ctx
Qwen2.5-Coder-32B-Instruct32.8B
Coding
DECENT35/100~2.9 tok/s
ctx
DECENT35/100~2.9 tok/s
ctx
Qwen2.5-Coder-32B-Instruct-AWQ32.8B
Coding
DECENT35/100~2.9 tok/s
ctx
DECENT35/100~2.9 tok/s
ctx
StarCoder2-15B15B
Coding
DECENT34/100~7.3 tok/s
ctx
DECENT34/100~7.3 tok/s
ctx
Qwen3-Coder-Next-AWQ-4bit14.4B
Coding
DECENT34/100~7.6 tok/s
ctx
DECENT34/100~7.6 tok/s
ctx
Qwen2.5-Coder-14B-Instruct14.8B
Coding
DECENT34/100~7.4 tok/s
ctx
DECENT34/100~7.4 tok/s
ctx
Qwen2.5-Coder-14B-Instruct-AWQ14.8B
Coding
DECENT34/100~7.4 tok/s
ctx
DECENT34/100~7.4 tok/s
ctx
Qwen3-Coder-30B-A3B-Instruct-FP415.6BMoE
Coding
DECENT34/100~5.6 tok/s
ctx
DECENT34/100~5.6 tok/s
ctx
DeepSeek-Coder-V2-Lite-Instruct15.7BMoE
Coding
DECENT34/100~5.6 tok/s
ctx
DECENT34/100~5.6 tok/s
ctx
DeepSeek-Coder-V2-Lite-Instruct-FP815.7BMoE
Coding
DECENT34/100~5.6 tok/s
ctx
DECENT34/100~5.6 tok/s
ctx
Qwen2.5-Coder-14B14B
Coding
DECENT33/100~7.9 tok/s
ctx
DECENT33/100~7.9 tok/s
ctx
CodeLlama-13B-Instruct13B
Coding
DECENT33/100~8.5 tok/s
ctx
DECENT33/100~8.5 tok/s
ctx
CodeLlama-13b-Instruct-hf13B
Coding
DECENT33/100~8.5 tok/s
ctx
DECENT33/100~8.5 tok/s
ctx
xLAM-8x7b-r46.7BMoE
General
DECENT32/100~1.5 tok/s
ctx
DECENT32/100~1.5 tok/s
ctx
Nous-Hermes-2-Mixtral-8x7B-DPO46.7BMoE
General
DECENT32/100~1.5 tok/s
ctx
DECENT32/100~1.5 tok/s
ctx
Llama-3_3-Nemotron-Super-49B-v1_549.9B
General
DECENT32/100~1.7 tok/s
ctx
DECENT32/100~1.7 tok/s
ctx
Llama-3_3-Nemotron-Super-49B-v1_5-FP849.9B
General
DECENT32/100~1.7 tok/s
ctx
DECENT32/100~1.7 tok/s
ctx
Llama-3_3-Nemotron-Super-49B-v149.9B
General
DECENT32/100~1.7 tok/s
ctx
DECENT32/100~1.7 tok/s
ctx
Mistral-Small-24B24B
General
DECENT31/100~4.3 tok/s
ctx
DECENT31/100~4.3 tok/s
ctx
Qwen2.5-32B-Instruct32B
General
DECENT31/100~2.9 tok/s
ctx
DECENT31/100~2.9 tok/s
ctx
gemma-2-27b-it27B
General
DECENT31/100~3.7 tok/s
ctx
DECENT31/100~3.7 tok/s
ctx
gemma-3-27b-it27B
General
DECENT31/100~3.7 tok/s
ctx
DECENT31/100~3.7 tok/s
ctx
Command-R35B
General
DECENT31/100~2.7 tok/s
ctx
DECENT31/100~2.7 tok/s
ctx
Falcon-40B-Instruct40B
General
DECENT31/100~2.1 tok/s
ctx
DECENT31/100~2.1 tok/s
ctx
t5gemma-9b-9b-ul220.3B
General
DECENT31/100~5.3 tok/s
ctx
DECENT30/100~5.3 tok/s
ctx
gpt-oss-20b21.5B
General
DECENT31/100~5 tok/s
ctx
DECENT31/100~5 tok/s
ctx
ERNIE-4.5-21B-A3B-MLX-4bit21.8BMoE
General
DECENT31/100~3.9 tok/s
ctx
DECENT31/100~3.9 tok/s
ctx
ERNIE-4.5-21B-A3B-MLX-6bit21.8BMoE
General
DECENT31/100~3.9 tok/s
ctx
DECENT31/100~3.9 tok/s
ctx
ERNIE-4.5-21B-A3B-MLX-8bit21.8BMoE
General
DECENT31/100~3.9 tok/s
ctx
DECENT31/100~3.9 tok/s
ctx
LFM2-24B-A2B-MLX-4bit23.8BMoE
General
DECENT31/100~3.5 tok/s
ctx
DECENT31/100~3.5 tok/s
ctx
LFM2-24B-A2B-MLX-6bit23.8BMoE
General
DECENT31/100~3.5 tok/s
ctx
DECENT31/100~3.5 tok/s
ctx
LFM2-24B-A2B-MLX-8bit23.8BMoE
General
DECENT31/100~3.5 tok/s
ctx
DECENT31/100~3.5 tok/s
ctx
LFM2-24B-A2B-MLX-5bit23.8BMoE
General
DECENT31/100~3.5 tok/s
ctx
DECENT31/100~3.5 tok/s
ctx
LFM2-24B-A2B23.8BMoE
General
DECENT31/100~3.5 tok/s
ctx
DECENT31/100~3.5 tok/s
ctx
Qwen3.5-27B27.8B
General
DECENT31/100~3.6 tok/s
ctx
DECENT31/100~3.6 tok/s
ctx
GLM-4.7-Flash-MLX-8bit29.9BMoE
General
DECENT31/100~2.6 tok/s
ctx
DECENT31/100~2.6 tok/s
ctx
GLM-4.7-Flash-MLX-6bit29.9BMoE
General
DECENT31/100~2.6 tok/s
ctx
DECENT31/100~2.6 tok/s
ctx
Qwen3-30B-A3B-Thinking-250730.5BMoE
General
DECENT31/100~2.5 tok/s
ctx
DECENT30/100~2.5 tok/s
ctx
Qwen3-30B-A3B-GPTQ-Int430.5BMoE
General
DECENT31/100~2.5 tok/s
ctx
DECENT30/100~2.5 tok/s
ctx
Qwen3-30B-A3B-Base30.5BMoE
General
DECENT31/100~2.5 tok/s
ctx
DECENT30/100~2.5 tok/s
ctx
Qwen3-30B-A3B-AWQ30.5BMoE
General
DECENT31/100~2.5 tok/s
ctx
DECENT30/100~2.5 tok/s
ctx
GLM-4.7-Flash31.2BMoE
General
DECENT31/100~2.4 tok/s
ctx
DECENT31/100~2.4 tok/s
ctx
GLM-4.7-Flash-AWQ31.2BMoE
General
DECENT31/100~2.4 tok/s
ctx
DECENT31/100~2.4 tok/s
ctx
NVIDIA-Nemotron-3-Nano-30B-A3B-MLX-4bit31.6B
General
DECENT31/100~3 tok/s
ctx
DECENT31/100~3 tok/s
ctx
NVIDIA-Nemotron-3-Nano-30B-A3B-MLX-8bit31.6B
General
DECENT31/100~3 tok/s
ctx
DECENT31/100~3 tok/s
ctx
NVIDIA-Nemotron-3-Nano-30B-A3B-MLX-6bit31.6B
General
DECENT31/100~3 tok/s
ctx
DECENT31/100~3 tok/s
ctx
NVIDIA-Nemotron-3-Nano-30B-A3B-MLX-5bit31.6B
General
DECENT31/100~3 tok/s
ctx
DECENT31/100~3 tok/s
ctx
NVIDIA-Nemotron-3-Nano-30B-A3B-BF1631.6B
General
DECENT31/100~3 tok/s
ctx
DECENT31/100~3 tok/s
ctx
NVIDIA-Nemotron-3-Nano-30B-A3B-Base-BF1631.6B
General
DECENT31/100~3 tok/s
ctx
DECENT31/100~3 tok/s
ctx
NVIDIA-Nemotron-3-Nano-30B-A3B-FP831.6B
General
DECENT31/100~3 tok/s
ctx
DECENT31/100~3 tok/s
ctx
EXAONE-4.0-32B32B
General
DECENT31/100~2.9 tok/s
ctx
DECENT31/100~2.9 tok/s
ctx
EXAONE-4.0.1-32B32B
General
DECENT31/100~2.9 tok/s
ctx
DECENT31/100~2.9 tok/s
ctx
EXAONE-4.0-32B-FP832B
General
DECENT31/100~2.9 tok/s
ctx
DECENT31/100~2.9 tok/s
ctx
sarvam-30b32.2BMoE
General
DECENT31/100~2.3 tok/s
ctx
DECENT31/100~2.3 tok/s
ctx
Olmo-3-1125-32B32.2B
General
DECENT31/100~2.9 tok/s
ctx
DECENT31/100~2.9 tok/s
ctx
Qwen3-32B-AWQ32.8B
General
DECENT31/100~2.9 tok/s
ctx
DECENT31/100~2.9 tok/s
ctx
Qwen2.5-32B32.8B
General
DECENT31/100~2.9 tok/s
ctx
DECENT31/100~2.9 tok/s
ctx
QwQ-32B-AWQ32.8B
General
DECENT31/100~2.9 tok/s
ctx
DECENT31/100~2.9 tok/s
ctx
Baichuan-M2-32B32.8B
General
DECENT31/100~2.9 tok/s
ctx
DECENT31/100~2.9 tok/s
ctx
QwQ-32B32.8B
General
DECENT31/100~2.9 tok/s
ctx
DECENT31/100~2.9 tok/s
ctx
xLAM-2-32b-fc-r32.8B
General
DECENT31/100~2.9 tok/s
ctx
DECENT31/100~2.9 tok/s
ctx
HyperCLOVAX-SEED-Think-32B33.3B
General
DECENT31/100~2.8 tok/s
ctx
DECENT31/100~2.8 tok/s
ctx
dolphin-2.9.1-yi-1.5-34b34.4B
General
DECENT31/100~2.7 tok/s
ctx
DECENT31/100~2.7 tok/s
ctx
Qwen3.5-35B-A3B36BMoE
General
DECENT31/100~2.1 tok/s
ctx
DECENT31/100~2.1 tok/s
ctx
Qwen3.8-27B27B
General
DECENT31/100~3.7 tok/s
ctx
DECENT31/100~3.7 tok/s
ctx
NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF1630B
General
DECENT31/100~3.3 tok/s
ctx
DECENT31/100~3.3 tok/s
ctx
Muse-Glimmer-30B30B
General
DECENT31/100~3.3 tok/s
ctx
DECENT31/100~3.3 tok/s
ctx
Laguna-XS-2.133B
General
DECENT31/100~2.8 tok/s
ctx
DECENT31/100~2.8 tok/s
ctx
North-Mini-Code-1.030B
General
DECENT31/100~3.3 tok/s
ctx
DECENT31/100~3.3 tok/s
ctx
Nemotron-3-Nano-Omni-30B-A3B-Reasoning-BF1630B
General
DECENT31/100~3.3 tok/s
ctx
DECENT31/100~3.3 tok/s
ctx
Qwen3.6-35B-A3B35B
General
DECENT31/100~2.7 tok/s
ctx
DECENT31/100~2.7 tok/s
ctx
Qwen3.6-27B27B
General
DECENT31/100~3.7 tok/s
ctx
DECENT31/100~3.7 tok/s
ctx
gemma-4-26B-A4B-it25.2B
General
DECENT31/100~4.1 tok/s
ctx
DECENT31/100~4.1 tok/s
ctx
gemma-4-31B-it31B
General
DECENT31/100~3 tok/s
ctx
DECENT31/100~3 tok/s
ctx
Voxtral-Small-24B-250724B
General
DECENT31/100~4.3 tok/s
ctx
DECENT31/100~4.3 tok/s
ctx
Qwen3-VL-32B-Instruct32B
General
DECENT31/100~2.9 tok/s
ctx
DECENT31/100~2.9 tok/s
ctx
Qwen3-VL-30B-A3B-Instruct30B
General
DECENT31/100~3.3 tok/s
ctx
DECENT31/100~3.3 tok/s
ctx
Qwen3-30B-A3B-Instruct-250730B
General
DECENT31/100~3.3 tok/s
ctx
DECENT31/100~3.3 tok/s
ctx
Mistral-Small-3.2-24B-Instruct-250624B
General
DECENT31/100~4.3 tok/s
ctx
DECENT31/100~4.3 tok/s
ctx
Mistral-Small-3.1-24B-Instruct-250324B
General
DECENT31/100~4.3 tok/s
ctx
DECENT31/100~4.3 tok/s
ctx
Mistral-Small-24B-Instruct-250124B
General
DECENT31/100~4.3 tok/s
ctx
DECENT31/100~4.3 tok/s
ctx
Bielik-11B-v3.0-Instruct11.2B
Chat
DECENT30/100~9.8 tok/s
ctx142
DECENT32/100~9.8 tok/s
142568
Qwen3-Next-80B-A3B-Thinking-AWQ-4bit14.7B
General
DECENT30/100~7.5 tok/s
ctx
DECENT30/100~7.5 tok/s
ctx
HyperCLOVAX-SEED-Think-14B-GPTQ14.7B
General
DECENT30/100~7.5 tok/s
ctx
DECENT30/100~7.5 tok/s
ctx
Qwen3-14B-AWQ14.8B
General
DECENT30/100~7.4 tok/s
ctx
DECENT30/100~7.4 tok/s
ctx
Qwen3-14B-Base14.8B
General
DECENT30/100~7.4 tok/s
ctx
DECENT30/100~7.4 tok/s
ctx
Qwen2.5-14B14.8B
General
DECENT30/100~7.4 tok/s
ctx
DECENT30/100~7.4 tok/s
ctx
Qwen3-30B-A3B-NVFP415.6BMoE
General
DECENT30/100~5.6 tok/s
ctx
DECENT30/100~5.6 tok/s
ctx
DeepSeek-V2-Lite15.7BMoE
General
DECENT30/100~5.6 tok/s
ctx
DECENT30/100~5.6 tok/s
ctx
Moonlight-16B-A3B16BMoE
General
DECENT30/100~5.5 tok/s
ctx
DECENT30/100~5.5 tok/s
ctx
deepseek-moe-16b-base16.4B
General
DECENT30/100~6.7 tok/s
ctx
DECENT30/100~6.7 tok/s
ctx
Qwen3-32B-NVFP417.2B
General
DECENT30/100~6.2 tok/s
ctx
DECENT30/100~6.2 tok/s
ctx
NVIDIA-Nemotron-3-Nano-30B-A3B-NVFP418.2B
General
DECENT30/100~5.9 tok/s
ctx
DECENT30/100~5.9 tok/s
ctx
Llama-4-Scout-17B-16E-Instruct17B
General
DECENT30/100~6.3 tok/s
ctx
DECENT30/100~6.3 tok/s
ctx
Mistral-Nemo-12B12B
General
DECENT29/100~9.2 tok/s
ctx
DECENT29/100~9.2 tok/s
ctx
Qwen2.5-14B-Instruct14B
General
DECENT29/100~7.9 tok/s
ctx
DECENT29/100~7.9 tok/s
ctx
gemma-3-12b-it12B
General
DECENT29/100~9.2 tok/s
ctx
DECENT29/100~9.2 tok/s
ctx
Phi-3-medium-128k14B
General
DECENT29/100~7.9 tok/s
ctx
DECENT29/100~7.9 tok/s
ctx
OLMo-2-13B-Instruct13B
General
DECENT29/100~8.5 tok/s
ctx
DECENT29/100~8.5 tok/s
ctx
pythia-12b12B
General
DECENT29/100~9.2 tok/s
ctx
DECENT29/100~9.2 tok/s
ctx
HarmBench-Llama-2-13b-cls13B
General
DECENT29/100~8.5 tok/s
ctx
DECENT29/100~8.5 tok/s
ctx
llm-jp-3.1-13b13.7B
General
DECENT29/100~8 tok/s
ctx
DECENT29/100~8 tok/s
ctx
Qwen1.5-MoE-A2.7B14.3BMoE
General
DECENT29/100~6.2 tok/s
ctx
DECENT29/100~6.2 tok/s
ctx
Ministral-3-14B-Instruct-251214B
General
DECENT29/100~7.9 tok/s
ctx
DECENT29/100~7.9 tok/s
ctx
Hunyuan-A13B-Instruct13B
General
DECENT29/100~8.5 tok/s
ctx
DECENT29/100~8.5 tok/s
ctx
Mistral-Nemo-Instruct-240712B
General
DECENT29/100~9.2 tok/s
ctx
DECENT29/100~9.2 tok/s
ctx
Seed-OSS-36B-Instruct-MLX-8bit36.2BMoE
Chat
DECENT23/100~2.1 tok/s
ctx
DECENT23/100~2.1 tok/s
ctx
Seed-OSS-36B-Instruct-MLX-4bit36.2BMoE
Chat
DECENT23/100~2.1 tok/s
ctx
DECENT23/100~2.1 tok/s
ctx
Seed-OSS-36B-Instruct-MLX-5bit36.2BMoE
Chat
DECENT23/100~2.1 tok/s
ctx
DECENT23/100~2.1 tok/s
ctx
Seed-OSS-36B-Instruct-MLX-6bit36.2BMoE
Chat
DECENT23/100~2.1 tok/s
ctx
DECENT23/100~2.1 tok/s
ctx
MiniMax-M2.5-AWQ-4bit36.8BMoE
Chat
DECENT23/100~2 tok/s
ctx
DECENT23/100~2 tok/s
ctx
Mixtral-8x7B-Instruct-v0.146.7BMoE
Chat
DECENT23/100~1.5 tok/s
ctx
DECENT23/100~1.5 tok/s
ctx
Kimi-Linear-48B-A3B-Instruct49.1B
Chat
DECENT23/100~1.7 tok/s
ctx
DECENT23/100~1.7 tok/s
ctx
llm-jp-3.1-13b-instruct413.7B
Chat
DECENT22/100~8 tok/s
ctx
DECENT22/100~8 tok/s
ctx
Qwen3-Next-80B-A3B-Instruct-AWQ-4bit14.7B
Chat
DECENT22/100~7.5 tok/s
ctx
DECENT22/100~7.5 tok/s
ctx
Qwen3-14B-Instruct14.8B
Chat
DECENT22/100~7.4 tok/s
ctx
DECENT22/100~7.4 tok/s
ctx
Qwen2.5-14B-Instruct-AWQ14.8B
Chat
DECENT22/100~7.4 tok/s
ctx
DECENT22/100~7.4 tok/s
ctx
Qwen2.5-14B-Instruct-GPTQ-Int414.8B
Chat
DECENT22/100~7.4 tok/s
ctx
DECENT22/100~7.4 tok/s
ctx
Qwen2.5-14B-Instruct-1M14.8B
Chat
DECENT22/100~7.4 tok/s
ctx
DECENT22/100~7.4 tok/s
ctx
Qwen2.5-14B-Instruct-GPTQ-Int814.8B
Chat
DECENT22/100~7.4 tok/s
ctx
DECENT22/100~7.4 tok/s
ctx
Qwen3-30B-A3B-Instruct-2507-FP415.6BMoE
Chat
DECENT22/100~5.6 tok/s
ctx
DECENT22/100~5.6 tok/s
ctx
DeepSeek-V2-Lite-Chat15.7BMoE
Chat
DECENT22/100~5.6 tok/s
ctx
DECENT22/100~5.6 tok/s
ctx
Moonlight-16B-A3B-Instruct16BMoE
Chat
DECENT22/100~5.5 tok/s
ctx
DECENT22/100~5.5 tok/s
ctx
LLaDA2.0-mini16.3BMoE
Chat
DECENT22/100~5.4 tok/s
ctx
DECENT22/100~5.4 tok/s
ctx
LLaDA2.1-mini16.3BMoE
Chat
DECENT22/100~5.4 tok/s
ctx
DECENT22/100~5.4 tok/s
ctx
deepseek-moe-16b-chat16.4B
Chat
DECENT22/100~6.7 tok/s
ctx
DECENT22/100~6.7 tok/s
ctx
Mistral-Small-24B-Instruct-2501-AWQ23.6B
Chat
DECENT22/100~4.4 tok/s
ctx
DECENT22/100~4.4 tok/s
ctx
Mistral-Small-24B-Instruct-2501-FP8-dynamic23.6B
Chat
DECENT22/100~4.4 tok/s
ctx
DECENT22/100~4.4 tok/s
ctx
Qwen3-30B-A3B-Instruct-2507-MLX-4bit30.5BMoE
Chat
DECENT22/100~2.5 tok/s
ctx
DECENT22/100~2.5 tok/s
ctx
Qwen3-30B-A3B-Instruct-2507-MLX-8bit30.5BMoE
Chat
DECENT22/100~2.5 tok/s
ctx
DECENT22/100~2.5 tok/s
ctx
Qwen3-30B-A3B-Instruct-2507-MLX-6bit30.5BMoE
Chat
DECENT22/100~2.5 tok/s
ctx
DECENT22/100~2.5 tok/s
ctx
Qwen3-30B-A3B-Instruct-2507-FP830.5BMoE
Chat
DECENT22/100~2.5 tok/s
ctx
DECENT22/100~2.5 tok/s
ctx
Qwen3-VL-30B-A3B-Instruct-AWQ31.1BMoE
Chat
DECENT22/100~2.4 tok/s
ctx
DECENT22/100~2.4 tok/s
ctx
OLMo-2-0325-32B-Instruct32.2B
Chat
DECENT22/100~2.9 tok/s
ctx
DECENT22/100~2.9 tok/s
ctx
Qwen2.5-32B-Instruct-AWQ32.8B
Chat
DECENT22/100~2.9 tok/s
ctx
DECENT22/100~2.9 tok/s
ctx
Qwen2.5-32B-Instruct-GPTQ-Int432.8B
Chat
DECENT22/100~2.9 tok/s
ctx
DECENT22/100~2.9 tok/s
ctx
Qwen2.5-32B-Instruct-GPTQ-Int832.8B
Chat
DECENT22/100~2.9 tok/s
ctx
DECENT22/100~2.9 tok/s
ctx
MiniMax-M2.5-BF16-INT4-AWQ39.1BMoE
Chat
DECENT22/100~1.8 tok/s
ctx
DECENT22/100~1.8 tok/s
ctx
GigaChat3-10B-A1.8B11.5BMoE
Chat
DECENT21/100~7.7 tok/s
ctx
DECENT21/100~7.7 tok/s
ctx
mistral-nemo-instruct-2407-awq12.2B
Chat
DECENT21/100~9 tok/s
ctx
DECENT22/100~9 tok/s
ctx
DeepSeek-R1-0528-NVFP4-v2393.6BMoE
Reasoning
TOO HEAVY5/100~0.3 tok/s
ctx
TOO HEAVY5/100~0.3 tok/s
ctx
DeepSeek-R1-NVFP4396.8BMoE
Reasoning
TOO HEAVY5/100~0.3 tok/s
ctx
TOO HEAVY5/100~0.3 tok/s
ctx
DeepSeek-R1684.5BMoE
Reasoning
TOO HEAVY5/100~0.2 tok/s
ctx
TOO HEAVY5/100~0.2 tok/s
ctx
DeepSeek-R1-0528684.5BMoE
Reasoning
TOO HEAVY5/100~0.2 tok/s
ctx
TOO HEAVY5/100~0.2 tok/s
ctx
DeepSeek-R1-70B70B
Reasoning
TOO HEAVY3/100~2.5 tok/s
ctx
TOO HEAVY3/100~2.5 tok/s
ctx
DeepSeek-R1-Distill-Llama-70B-FP8-dynamic70.6B
Reasoning
TOO HEAVY3/100~2.4 tok/s
ctx
TOO HEAVY3/100~2.4 tok/s
ctx
Llama-3.1-70B-Instruct70B
General
TOO HEAVY0/100~2.5 tok/s
ctx
TOO HEAVY0/100~2.5 tok/s
ctx
Llama-3.1-405B-Instruct405B
General
TOO HEAVY0/100~0.4 tok/s
ctx
TOO HEAVY0/100~0.4 tok/s
ctx
Llama-3.2-90B-Vision90B
Multimodal
TOO HEAVY0/100~1.9 tok/s
ctx
TOO HEAVY0/100~1.9 tok/s
ctx
Mixtral-8x22B-Instruct141BMoE
General
TOO HEAVY0/100~1 tok/s
ctx
TOO HEAVY0/100~1 tok/s
ctx
Qwen2.5-72B-Instruct72B
General
TOO HEAVY0/100~2.4 tok/s
ctx
TOO HEAVY0/100~2.4 tok/s
ctx
Qwen2.5-VL-72B72B
Multimodal
TOO HEAVY0/100~2.4 tok/s
ctx
TOO HEAVY0/100~2.4 tok/s
ctx
DeepSeek-V3671BMoE
General
TOO HEAVY0/100~0.2 tok/s
ctx
TOO HEAVY0/100~0.2 tok/s
ctx
Command-R+104B
General
TOO HEAVY0/100~1.7 tok/s
ctx
TOO HEAVY0/100~1.7 tok/s
ctx
Qwen3.5-122B-A10B-NVFP464.4BMoE
General
TOO HEAVY0/100~2.1 tok/s
ctx
TOO HEAVY0/100~2.1 tok/s
ctx
NVIDIA-Nemotron-3-Super-120B-A12B-NVFP467.2B
General
TOO HEAVY0/100~2.6 tok/s
ctx
TOO HEAVY0/100~2.6 tok/s
ctx
Llama-3.3-70B-Instruct70.6B
Chat
TOO HEAVY0/100~2.4 tok/s
ctx
TOO HEAVY0/100~2.4 tok/s
ctx
llama-3.3-70b-instruct-awq70.6B
Chat
TOO HEAVY0/100~2.4 tok/s
ctx
TOO HEAVY0/100~2.4 tok/s
ctx
Llama-3.3-70B-Instruct-AWQ70.6B
Chat
TOO HEAVY0/100~2.4 tok/s
ctx
TOO HEAVY0/100~2.4 tok/s
ctx
L3.3-GeneticLemonade-Final-v2-70B70.6B
General
TOO HEAVY0/100~2.4 tok/s
ctx
TOO HEAVY0/100~2.4 tok/s
ctx
Meta-Llama-3.3-70B-Instruct-AWQ-INT470.6B
Chat
TOO HEAVY0/100~2.4 tok/s
ctx
TOO HEAVY0/100~2.4 tok/s
ctx
Meta-Llama-3.1-70B-Instruct-quantized.w4a1670.6B
Chat
TOO HEAVY0/100~2.4 tok/s
ctx
TOO HEAVY0/100~2.4 tok/s
ctx
Meta-Llama-3-70B-Instruct70.6B
Chat
TOO HEAVY0/100~2.4 tok/s
ctx
TOO HEAVY0/100~2.4 tok/s
ctx
Llama-3.1-70B70.6B
General
TOO HEAVY0/100~2.4 tok/s
ctx
TOO HEAVY0/100~2.4 tok/s
ctx
Llama-3.1-Swallow-70B-Instruct-v0.370.6B
Chat
TOO HEAVY0/100~2.4 tok/s
ctx
TOO HEAVY0/100~2.4 tok/s
ctx
Meta-Llama-3.1-70B-Instruct-FP870.6B
Chat
TOO HEAVY0/100~2.4 tok/s
ctx
TOO HEAVY0/100~2.4 tok/s
ctx
Llama-3.3-70B-Instruct-FP8-dynamic70.6B
Chat
TOO HEAVY0/100~2.4 tok/s
ctx
TOO HEAVY0/100~2.4 tok/s
ctx
jais-adapted-70b-chat-4bit-bnb71.6B
Chat
TOO HEAVY0/100~2.4 tok/s
ctx
TOO HEAVY0/100~2.4 tok/s
ctx
Qwen2.5-72B-Instruct-abliterated72.7B
Chat
TOO HEAVY0/100~2.4 tok/s
ctx
TOO HEAVY0/100~2.4 tok/s
ctx
Qwen2.5-72B72.7B
General
TOO HEAVY0/100~2.4 tok/s
ctx
TOO HEAVY0/100~2.4 tok/s
ctx
Qwen2-72B-Instruct72.7B
Chat
TOO HEAVY0/100~2.4 tok/s
ctx
TOO HEAVY0/100~2.4 tok/s
ctx
Qwen2-72B72.7B
General
TOO HEAVY0/100~2.4 tok/s
ctx
TOO HEAVY0/100~2.4 tok/s
ctx
Qwen2.5-72B-Instruct-AWQ73B
Chat
TOO HEAVY0/100~2.4 tok/s
ctx
TOO HEAVY0/100~2.4 tok/s
ctx
Qwen2.5-72B-Instruct-GPTQ-Int473B
Chat
TOO HEAVY0/100~2.4 tok/s
ctx
TOO HEAVY0/100~2.4 tok/s
ctx
Qwen3-Coder-Next-8bit79.7B
Coding
TOO HEAVY0/100~2.2 tok/s
ctx
TOO HEAVY0/100~2.2 tok/s
ctx
Qwen3-Next-80B-A3B-Instruct-MLX-4bit79.7B
Chat
TOO HEAVY0/100~2.2 tok/s
ctx
TOO HEAVY0/100~2.2 tok/s
ctx
Qwen3-Next-80B-A3B-Instruct-MLX-8bit79.7B
Chat
TOO HEAVY0/100~2.2 tok/s
ctx
TOO HEAVY0/100~2.2 tok/s
ctx
Qwen3-Next-80B-A3B-Instruct-MLX-6bit79.7B
Chat
TOO HEAVY0/100~2.2 tok/s
ctx
TOO HEAVY0/100~2.2 tok/s
ctx
Qwen3-Next-80B-A3B-Instruct-MLX-5bit79.7B
Chat
TOO HEAVY0/100~2.2 tok/s
ctx
TOO HEAVY0/100~2.2 tok/s
ctx
Qwen3-Coder-Next79.7B
Coding
TOO HEAVY0/100~2.2 tok/s
ctx
TOO HEAVY0/100~2.2 tok/s
ctx
Qwen3-Coder-Next-FP879.7B
Coding
TOO HEAVY0/100~2.2 tok/s
ctx
TOO HEAVY0/100~2.2 tok/s
ctx
Qwen3-Next-80B-A3B-Instruct81.3B
Chat
TOO HEAVY0/100~2.1 tok/s
ctx
TOO HEAVY0/100~2.1 tok/s
ctx
Qwen3-Next-80B-A3B-Instruct-FP881.3B
Chat
TOO HEAVY0/100~2.1 tok/s
ctx
TOO HEAVY0/100~2.1 tok/s
ctx
GLM-4.5-Air110.5BMoE
General
TOO HEAVY0/100~1.2 tok/s
ctx
TOO HEAVY0/100~1.2 tok/s
ctx
Qwen1.5-110B-Chat-AWQ111.2B
Chat
TOO HEAVY0/100~1.5 tok/s
ctx
TOO HEAVY0/100~1.5 tok/s
ctx
gpt-oss-120b-MLX-8bit116.8B
General
TOO HEAVY0/100~1.5 tok/s
ctx
TOO HEAVY0/100~1.5 tok/s
ctx
gpt-oss-120b-heretic116.8B
General
TOO HEAVY0/100~1.5 tok/s
ctx
TOO HEAVY0/100~1.5 tok/s
ctx
gpt-oss-120b120.4B
General
TOO HEAVY0/100~1.4 tok/s
ctx
TOO HEAVY0/100~1.4 tok/s
ctx
XORTRON.CriminalComputing.LARGE.2026.3122.6B
General
TOO HEAVY0/100~1.4 tok/s
ctx
TOO HEAVY0/100~1.4 tok/s
ctx
NVIDIA-Nemotron-3-Super-120B-A12B-FP8123.6B
General
TOO HEAVY0/100~1.4 tok/s
ctx
TOO HEAVY0/100~1.4 tok/s
ctx
NVIDIA-Nemotron-3-Super-120B-A12B-BF16123.6B
General
TOO HEAVY0/100~1.4 tok/s
ctx
TOO HEAVY0/100~1.4 tok/s
ctx
Qwen3.5-122B-A10B125.1BMoE
General
TOO HEAVY0/100~1.1 tok/s
ctx
TOO HEAVY0/100~1.1 tok/s
ctx
Mixtral-8x22B-Instruct-v0.1140.6BMoE
Chat
TOO HEAVY0/100~1 tok/s
ctx
TOO HEAVY0/100~1 tok/s
ctx
dots.llm1.inst142.8BMoE
General
TOO HEAVY0/100~1 tok/s
ctx
TOO HEAVY0/100~1 tok/s
ctx
bloom176.2B
General
TOO HEAVY0/100~1 tok/s
ctx
TOO HEAVY0/100~1 tok/s
ctx
GLM-4.7-NVFP4177.2BMoE
General
TOO HEAVY0/100~0.8 tok/s
ctx
TOO HEAVY0/100~0.8 tok/s
ctx
Step-3.5-Flash199.4BMoE
General
TOO HEAVY0/100~0.7 tok/s
ctx
TOO HEAVY0/100~0.7 tok/s
ctx
Step-3.5-Flash-FP8199.4BMoE
General
TOO HEAVY0/100~0.7 tok/s
ctx
TOO HEAVY0/100~0.7 tok/s
ctx
MiniMax-M2.5-MLX-8bit228.7BMoE
Chat
TOO HEAVY0/100~0.6 tok/s
ctx
TOO HEAVY0/100~0.6 tok/s
ctx
MiniMax-M2.5-MLX-4bit228.7BMoE
Chat
TOO HEAVY0/100~0.6 tok/s
ctx
TOO HEAVY0/100~0.6 tok/s
ctx
MiniMax-M2.5-MLX-6bit228.7BMoE
Chat
TOO HEAVY0/100~0.6 tok/s
ctx
TOO HEAVY0/100~0.6 tok/s
ctx
MiniMax-M2-AWQ228.7BMoE
Chat
TOO HEAVY0/100~0.6 tok/s
ctx
TOO HEAVY0/100~0.6 tok/s
ctx
MiniMax-M2.5-AWQ228.7BMoE
Chat
TOO HEAVY0/100~0.6 tok/s
ctx
TOO HEAVY0/100~0.6 tok/s
ctx
MiniMax-M2.5228.7BMoE
Chat
TOO HEAVY0/100~0.6 tok/s
ctx
TOO HEAVY0/100~0.6 tok/s
ctx
MiniMax-M2228.7BMoE
Chat
TOO HEAVY0/100~0.6 tok/s
ctx
TOO HEAVY0/100~0.6 tok/s
ctx
MiniMax-M2.1228.7BMoE
Chat
TOO HEAVY0/100~0.6 tok/s
ctx
TOO HEAVY0/100~0.6 tok/s
ctx
Qwen3-235B-A22B-Instruct-2507-FP8235.1BMoE
Chat
TOO HEAVY0/100~0.6 tok/s
ctx
TOO HEAVY0/100~0.6 tok/s
ctx
Qwen3-235B-A22B-Thinking-2507-FP8235.1BMoE
General
TOO HEAVY0/100~0.6 tok/s
ctx
TOO HEAVY0/100~0.6 tok/s
ctx
Qwen3-235B-A22B-FP8235.1BMoE
General
TOO HEAVY0/100~0.6 tok/s
ctx
TOO HEAVY0/100~0.6 tok/s
ctx
deepseek-coder-v2-instruct-awq235.7BMoE
Coding
TOO HEAVY0/100~0.6 tok/s
ctx
TOO HEAVY0/100~0.6 tok/s
ctx
DeepSeek-V2.5-1210-FP8235.7BMoE
General
TOO HEAVY0/100~0.6 tok/s
ctx
TOO HEAVY0/100~0.6 tok/s
ctx
K-EXAONE-236B-A23B237.1BMoE
General
TOO HEAVY0/100~0.6 tok/s
ctx
TOO HEAVY0/100~0.6 tok/s
ctx
MiMo-V2-Flash309.8BMoE
General
TOO HEAVY0/100~0.4 tok/s
ctx
TOO HEAVY0/100~0.4 tok/s
ctx
GLM-4.6356.8BMoE
General
TOO HEAVY0/100~0.4 tok/s
ctx
TOO HEAVY0/100~0.4 tok/s
ctx
GLM-4.7358.3BMoE
General
TOO HEAVY0/100~0.4 tok/s
ctx
TOO HEAVY0/100~0.4 tok/s
ctx
GLM-4.5358.3BMoE
General
TOO HEAVY0/100~0.4 tok/s
ctx
TOO HEAVY0/100~0.4 tok/s
ctx
DeepSeek-V3.2-NVFP4394.5BMoE
General
TOO HEAVY0/100~0.3 tok/s
ctx
TOO HEAVY0/100~0.3 tok/s
ctx
DeepSeek-V3-0324-NVFP4396.8BMoE
General
TOO HEAVY0/100~0.3 tok/s
ctx
TOO HEAVY0/100~0.3 tok/s
ctx
Llama-4-Maverick-17B-128E-Instruct401.6B
Chat
TOO HEAVY0/100~0.4 tok/s
ctx
TOO HEAVY0/100~0.4 tok/s
ctx
Qwen3.5-397B-A17B403.4BMoE
General
TOO HEAVY0/100~0.3 tok/s
ctx
TOO HEAVY0/100~0.3 tok/s
ctx
Llama-3.1-405B405.9B
General
TOO HEAVY0/100~0.4 tok/s
ctx
TOO HEAVY0/100~0.4 tok/s
ctx
Qwen3-Coder-480B-A35B-Instruct480.2BMoE
Coding
TOO HEAVY0/100~0.3 tok/s
ctx
TOO HEAVY0/100~0.3 tok/s
ctx
LongCat-Flash-Chat561.9B
Chat
TOO HEAVY0/100~0.3 tok/s
ctx
TOO HEAVY0/100~0.3 tok/s
ctx
DeepSeek-V3-0324684.5BMoE
General
TOO HEAVY0/100~0.2 tok/s
ctx
TOO HEAVY0/100~0.2 tok/s
ctx
DeepSeek-V3.2-AWQ685BMoE
General
TOO HEAVY0/100~0.2 tok/s
ctx
TOO HEAVY0/100~0.2 tok/s
ctx
DeepSeek-V3.2685.4BMoE
General
TOO HEAVY0/100~0.2 tok/s
ctx
TOO HEAVY0/100~0.2 tok/s
ctx
GLM-5753.9BMoE
General
TOO HEAVY0/100~0.2 tok/s
ctx
TOO HEAVY0/100~0.2 tok/s
ctx
GLM-5-FP8753.9BMoE
General
TOO HEAVY0/100~0.2 tok/s
ctx
TOO HEAVY0/100~0.2 tok/s
ctx
Kimi-K2-Instruct1026.5BMoE
Chat
TOO HEAVY0/100~0.1 tok/s
ctx
TOO HEAVY0/100~0.1 tok/s
ctx
Kimi-K2-Instruct-09051026.5BMoE
Chat
TOO HEAVY0/100~0.1 tok/s
ctx
TOO HEAVY0/100~0.1 tok/s
ctx
Kimi-K2-Thinking1058.1BMoE
General
TOO HEAVY0/100~0.1 tok/s
ctx
TOO HEAVY0/100~0.1 tok/s
ctx
Kimi-K2.51058.6BMoE
General
TOO HEAVY0/100~0.1 tok/s
ctx
TOO HEAVY0/100~0.1 tok/s
ctx
Qwen3.8-2.4T-A95B2400B
General
TOO HEAVY0/100~0.1 tok/s
ctx
TOO HEAVY0/100~0.1 tok/s
ctx
DeepSeek-V4-Flash-0731284B
General
TOO HEAVY0/100~0.6 tok/s
ctx
TOO HEAVY0/100~0.6 tok/s
ctx
Inkling-Small276B
General
TOO HEAVY0/100~0.6 tok/s
ctx
TOO HEAVY0/100~0.6 tok/s
ctx
Ling-3.0-flash124B
General
TOO HEAVY0/100~1.4 tok/s
ctx
TOO HEAVY0/100~1.4 tok/s
ctx
Inkling975B
General
TOO HEAVY0/100~0.2 tok/s
ctx
TOO HEAVY0/100~0.2 tok/s
ctx
Kimi-K32800B
General
TOO HEAVY0/100~0.1 tok/s
ctx
TOO HEAVY0/100~0.1 tok/s
ctx
Nex-N2-Pro397B
General
TOO HEAVY0/100~0.4 tok/s
ctx
TOO HEAVY0/100~0.4 tok/s
ctx
NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16550B
General
TOO HEAVY0/100~0.3 tok/s
ctx
TOO HEAVY0/100~0.3 tok/s
ctx
DeepSeek-V4-Pro1600B
General
TOO HEAVY0/100~0.1 tok/s
ctx
TOO HEAVY0/100~0.1 tok/s
ctx
DeepSeek-V4-Flash284B
General
TOO HEAVY0/100~0.6 tok/s
ctx
TOO HEAVY0/100~0.6 tok/s
ctx
RUNS GREAT
RUNS WELL
DECENT
TOO HEAVY
✦ TurboQuant — 4× KV compression · 4-bit · ~99.5% fidelity

What this comparison actually measures

Two numbers decide whether a local model is usable on a given machine, and this page compares both. Memory capacity decides whether the model loads at all. Memory bandwidth decides how fast it generates once it has, because producing a single token means reading every active weight once.

Compute throughput barely enters into single-stream decoding, which is why two cards with similar TFLOPs can feel completely different, and why an older card with more memory often beats a newer one with less.

How fit and speed are calculated

Every figure on this page comes from the same arithmetic used across the rest of the site, so a verdict here agrees with the calculators and the hardware pages.

  • Weights are parameters multiplied by bits per weight, divided by eight. A 27B model at Q4_K_M is roughly 15 GB.
  • KV cache is added on top and grows with context length. It is the term that turns a model which loaded cleanly into an out-of-memory error an hour later.
  • Runtime overhead of about a gigabyte covers the CUDA or Metal context and activation buffers.
  • Usable memory is roughly 90% of nameplate on a discrete card and lower on unified memory, where the operating system shares the pool.
  • Throughput is memory bandwidth divided by the bytes read per token, at about 80% efficiency.

These are engineering estimates, not benchmark results. Quantisation, runtime and context length all move the real number. See the methodology for every assumption behind them.

Comparing two devices: common questions

Why does an older GPU sometimes beat a newer one?

Because it has more memory. A previous-generation card with 24 GB runs models a newer 12 GB card cannot load at all, and a model that does not fit does not run slowly — it does not run. Capacity is the harder constraint, so it is the first thing to compare.

Should I compare on VRAM or on bandwidth?

Capacity first, bandwidth second. Work out which models each device can hold, and only then compare speed on the models both can run. Comparing tokens per second between two devices that fit different models is comparing different products.

How does unified memory compare to a discrete card?

Apple Silicon shares one pool between CPU, GPU and the system, so capacity is cheap and bandwidth is what separates the chip tiers. A high-end unified machine holds far larger models than a consumer card at similar money, and runs them more slowly. Which one wins depends entirely on the size of model you need.

What does the TurboQuant toggle change?

It applies a quantisation profile to device B only, so you can see what dropping precision buys in fit and context on one side of the comparison while the other stays fixed. Lower precision means smaller weights, which means both more headroom and faster generation — at some cost in quality.

Does more system RAM help if the model does not fit in VRAM?

It lets a larger model load, but it does not make it fast. Layers that spill to system RAM run at a fraction of card bandwidth, and because every token passes through every layer, the slow side dominates. A lower quantisation that fits entirely on the card is usually faster than offloading.

Work from a single number instead

If you already know which model you want, it is usually quicker to start from the model and let the arithmetic name the hardware: