This year's open models
- MiMo-V2.6-Pro-RLXiaomi (MiMo) · 2026-09Ranked #1 open-weights model on Artificial Analysis1.02T · 42B active
- DeepSeek-V4.1-FlashDeepSeek · 2026-09KV cache about 4x smaller than V4-Flash552B · 16B active
- GLM-5.3Z.ai (Zhipu) · 2026-08Top open coding and agent model744B · 40B active
- Qwen3.8-27BAlibaba (Qwen) · 2026-08Single-GPU dense model for coding and agents27B dense
- Muse-Glimmer-30BMeta (Superintelligence Labs) · 2026-08First Meta open weights since Llama 429.6B dense
- Kimi-K3Moonshot AI · 2026-07Frontier open model. Datacenter only.2.8T · 104B active
- Qwen3.6-35B-A3BAlibaba (Qwen) · 2026-04Agentic coding upgrade of Qwen3.5-35B-A3B35B · 3B active
- gemma-4-31BGoogle DeepMind · 2026-03Largest dense Gemma 4. General and multilingual.30.7B dense
How we estimate
Memory needed = weights + KV cache + 1 GB overhead. It must fit in the usable memory of the machine.
- Weights (GB) = total parameters (B) × bits per weight ÷ 8. Bits: Q4_K_M 4.85, Q8_0 8.5, BF16 16, MXFP4 4.25.
- KV cache (GB) = KV bytes per token × context, plus the sliding-window cap if the model has one. The cache uses FP16.
- If the KV size is unknown, the cell counts the weights only and shows *.
- Decode speed (tok/s) ≈ 0.7 × bandwidth ÷ (active weights in GB + KV GB read per token). The KV read counts half the context.
Usable memory is an estimate.
- Mac: the GPU may use about 2/3 of the memory up to 32 GB, and about 3/4 above that. Apple does not publish this limit.
- Mac with the limit raised: memory − 4 GB.
- DGX: memory − 8 GB for the OS.
- GeForce, RTX PRO and datacenter GPUs: VRAM × 0.95.
A discrete GPU can spill. The weights that do not fit stay in 64 GB of DDR5 at 89.6 GB/s. Per-token time is the VRAM share at GPU bandwidth plus the RAM share at 89.6 GB/s. A Mac or a DGX never spills.
These numbers are estimates, not benchmarks. One weight format applies to every model. Real runs depend on the runtime.
Learn more: Chapter 3, bandwidth · Chapter 4, will it fit · Chapter 10, MoE.
Sources: Apple Mac mini and Mac Studio specs; NVIDIA product pages; Hugging Face model cards and config.json files; llama.cpp discussion 2182 for the Mac GPU limit. Data as of 2026-10-11.