GLM-5.2 vs DeepSeek V4 vs Kimi K3: Which Giant MoE Fits Your Hardware

GLM-5.2, DeepSeek V4, and Kimi K3 have each been called the best open-weight model of 2026, but the leaderboard scores skip over a much more practical question: what does it actually take to load one of these on hardware you own? GLM-5.2's own 2-bit quant still needs roughly 239GB of memory, DeepSeek V4-Flash fits in about 128GB thanks to a much lighter active-expert count, and Kimi K3 has no GGUF at all as of this week. Here's what the numbers actually say about which one is worth your download bandwidth.

All three of these releases are sparse mixture-of-experts (MoE) models, not dense transformers, and that distinction is the whole story. DeepSeek V4-Flash has 284 billion total parameters but only about 13 billion activate for any given token. Kimi K3 is more extreme: roughly 2.8 trillion total parameters spread across 896 experts, of which only 16 fire per token, putting active compute around 50 billion parameters. GLM-5.2 sits in between at 744 billion total with about 40 billion active.

Active parameters are what determine how fast the model runs — closer to a mid-sized dense model than the eye-watering total would suggest. What they don't determine is memory footprint. Every expert has to sit somewhere the inference engine can reach, because the router can send the next token to any of them. Total parameter count, not active parameter count, is the number that decides whether a quantized build fits your VRAM and RAM. Our explainer on [how local LLM inference actually works](/blog/how-local-llm-inference-actually-works) covers why loading, not raw compute, is usually the first wall you hit.

It's worth noticing how fast the ceiling has moved. A year ago, a 70B dense model was considered a big local build, and 24GB of VRAM covered most of what a serious hobbyist needed to worry about. These three releases are one to two orders of magnitude larger in total parameters, and none of that growth came from making inference faster — it came from adding more experts that mostly sit idle for any single token. That's a deliberate tradeoff by the labs training these models: more total capacity and specialization, at the cost of a memory bill that now rivals a small server rack.

Quantization is what makes any of this conceivable on hardware short of a datacenter rack. Unsloth's 2-bit dynamic quant for GLM-5.2 (UD-IQ2_XXS) compresses the model from 1.51TB down to about 239GB — still enormous, but small enough to spread across a multi-GPU rig plus system RAM. DeepSeek V4-Flash's lighter expert count means the independent ds4 inference engine runs it inside 128GB of unified memory on a single Mac. Kimi K3 has no official or community GGUF yet; scaling from its predecessor Kimi K2 suggests even a punishing quant would land somewhere between 650GB and 1TB.

One number worth sitting with: Apple pulled the 256GB and 512GB memory options from the Mac Studio lineup in 2026 as RAM prices climbed, so the current M3 Ultra tops out at 96GB. That's the same ceiling as four RTX 3090s' combined VRAM — meaning neither a maxed-out current Mac Studio nor a 4-GPU 3090 rig can hold GLM-5.2's 239GB footprint without spilling a large chunk to slower memory.

Sort of, and slowly. Four RTX 3090s (used units currently run $700-$1,250 each, so roughly $3,000-$5,000 for the set) provide 96GB of VRAM. Add 192GB of system RAM and llama.cpp can keep GLM-5.2's always-resident layers and active expert path on the GPUs while streaming the inactive experts in from RAM as the router selects them. Community reports put realistic throughput at 3-9 tokens per second on that kind of setup — usable for asynchronous work, not for a snappy chat interface.

There's a real caveat here too: llama.cpp currently runs GLM-5.2 with a dense-attention fallback, because the model's sparse-attention path isn't implemented in mainline yet. That means long-context prompts are slower and heavier than GLM-5.2's architecture would suggest on paper — worth knowing before you plan a build around a big context window. If you're still sizing a GPU purchase, our [best GPU for local LLMs](/blog/best-gpu-for-local-llms-2026) guide and [VRAM sizing walkthrough](/blog/how-much-vram-to-run-local-llms) cover the more common single-GPU case.

Run the cost math before you buy the fourth card. A $3,000-$5,000 GPU rig plus a 192GB-capable motherboard and CPU is closer to $4,500-$6,500 all-in, dedicated to a model that outputs single-digit tokens per second. Renting an equivalent multi-GPU node by the hour is almost certainly cheaper if you only need GLM-5.2 occasionally — the rig only pays for itself if you're running it daily for months. That tradeoff is exactly what our [local LLM vs OpenAI API cost comparison](/blog/local-llm-vs-openai-api-cost) walks through for smaller models, and the same logic scales up here.

DeepSeek shipped two variants: V4-Pro at 1.6 trillion total parameters with 49 billion active, and V4-Flash at 284 billion total with 13 billion active. V4-Pro has no documented consumer-quantized footprint yet and is squarely a datacenter model for now. V4-Flash is the one people are actually running at home — the same engineer behind Redis built ds4, a purpose-built inference engine that loads V4-Flash on a single 128GB MacBook.

The gap between V4-Flash and the other two comes down to total parameter count, not cleverness in the quantization. 284B total is roughly a third of GLM-5.2's 744B and a tenth of Kimi K3's 2.8T, so even without an exotic 2-bit quant, V4-Flash's footprint stays inside what a single high-memory machine can hold.

Kimi K3 is Moonshot's follow-up to K2, and at 2.8 trillion total parameters across 896 experts (16 active per token, roughly 50B active) it's the largest open-weight release to date — we covered the [full architecture and benchmark story](/blog/kimi-k3-largest-open-weight-model-run-locally) when it launched. As of July 20, 2026, no GGUF, Ollama, or MLX build exists; the native MXFP4 weights alone are about 1.4TB.

Scaling from Kimi K2, which needed at least 128GB of unified memory just for its smallest quants, K3's quantized builds will likely land between roughly 650GB and 1TB even at aggressive settings — comfortably outside anything a consumer or prosumer rig can hold today. Treat K3 as a benchmark story for now, not a download.

Once a community GGUF exists for any of these models, the same three llama.cpp flags apply. `--cpu-moe` is the blunt instrument — it keeps every mixture-of-experts weight tensor in system RAM and runs everything else on GPU, which is the easiest way to get an oversized MoE model loading at all. `--n-cpu-moe N` offloads only the first N expert layers to CPU, useful once you know how many layers your VRAM can hold. `--override-tensor` (`-ot`) gives you regex control over exactly which tensors land where, for squeezing out the last few GB of headroom.

# Simplest: keep every MoE expert tensor in system RAM, run the rest on GPU
llama-server -m ./GLM-5.2-UD-IQ2_XXS.gguf -ngl 999 --cpu-moe -fa --ctx-size 4096

# Offload only the first N expert layers to CPU once you know your VRAM headroom
llama-server -m ./GLM-5.2-UD-IQ2_XXS.gguf -ngl 999 --n-cpu-moe 60 -fa

# Fine-grained: pin later-layer experts in VRAM, spill earlier layers to system RAM
llama-server -m ./GLM-5.2-UD-IQ2_XXS.gguf -ngl 999 \
  -ot "blk\.(0|[1-9]|[1-8][0-9])\.ffn_.*_exps\.weight=CPU" -fa

Finding the right `--n-cpu-moe` value for your card is trial and error — start high, drop it until you run out of VRAM, and expect the sweet spot to be specific to your exact GPU and quant.

If you have to pick one this month, match the model to what you actually own:

• ▸GLM-5.2 (Z.ai/Zhipu): 744B total / 40B active — 91.2% GPQA Diamond, 62.1% SWE-bench Pro. Needs ~239GB at 2-bit quantization. Realistic only on a 4x RTX 3090 rig with 192GB system RAM, or a maxed prior-generation Mac Studio with 256GB+ memory.
• ▸DeepSeek V4-Flash: 284B total / 13B active. Runs today inside 128GB of unified memory via the community ds4 engine — the only one of the three that's genuinely usable on a single high-end workstation.
• ▸DeepSeek V4-Pro: 1.6T total / 49B active. No documented consumer quantization path yet; datacenter-only for now.
• ▸Kimi K3: 2.8T total / ~50B active across 16 of 896 experts. No GGUF as of July 20, 2026; estimated 650GB-1TB even quantized. Read the full local-hardware reality check before you wait on this one.

For most Runyard readers, that means DeepSeek V4-Flash is the only one of the three worth downloading this week — the other two are either a multi-GPU project or, in Kimi K3's case, not yet possible outside a rented cluster. If quantization tradeoffs are new to you, our breakdown of [Q4_K_M vs Q5_K_M vs Q6_K vs Q8_0](/blog/q4-k-m-vs-q5-k-m-vs-q6-k-vs-q8-0-quantization-quality-loss) is worth reading before you pick a quant for any of them.

Before you commit a weekend to downloading a 239GB quant, check what your own hardware can actually hold. The [Runyard VRAM Calculator](/tools/vram-calculator) estimates which quant of GLM-5.2, DeepSeek V4-Flash, or any other model fits your VRAM, system RAM, and target context length. [Runyard Compare](/compare) puts these three head to head against smaller, more forgiving alternatives, and [Model Radar](/) surfaces the best model your exact setup can run today.