Compatibility Matrix

Best community run per VRAM tier for the most-run open-weight models. Cells show tok/s and quantization — the same model often lands in multiple tiers via different quants. Full board · Submit your run.

Model familyAll Arcee Baichuan BTL Cohere Deepseek Deepseek-Coder Deepseek-R1 DeepSeek-V4.1-Flash Gemma GLM Gpt Hunyuan Kimi Kwaipilot Laguna LFM Ling Llama MiniCPM-V Minimax Mistral MobileLLM Muse Nemotron Ornith Phi Qwen Qwen3.8 RNJ Spark unknown

Model	Family	<24GB	24–48GB	>48GB	runs
LFM2.5-2.6B	Llama	—	305.6 · Q4_K_M	—	2
MiniCPM5-1B	Llama	158.1 · Q4_K_M	—	—	2
granite-4.2-8b	Llama	52.5 · Q5_K_M	—	—	1
SmolLM2-135M-Instruct	Llama	—	1102.1 · Q2_K	—	1
Llama-3.2-3B-Instruct	Llama	—	—	489.2 · Q4_K_M	1
SmolLM2-360M-Instruct	Llama	—	—	968.1 · Q8_0	1
Llama-3.2-1B-Instruct	Llama	—	—	699.9 · Q8_0	1
Falcon3-7B-Instruct	Llama	—	19.9 · bfp_bf8	—	1
Llama-3.3-70B-Instruct	Llama	—	—	20.1 · NVFP4	1
gemma-4-12B-coder-fable5-composer2.5-v1	Llama	—	85.3 · Q8_0	—	1
LFM2-24B-A2B	Llama	—	393.2 · Q4_K_M	—	1
granite-4.2-30b	Llama	—	—	23.7 · Q8_0	1
Llama-3.1-8B-Instruct	Llama	—	—	158.1 · FP8	1
BTL-4-Compact	Llama	58.5 · IQ2_XXS	—	—	1
LFM2.5-1.2B-Instruct	Llama	—	709.5 · Q4_K_M	—	1
Meta-Llama-3-8B-Instruct	Llama	148.3 · Q4_0	—	—	1
reka-flash-3	Llama	23.9 · Q4_K_M	—	—	1
Llama-4-Scout-17B-16E-Instruct	Llama	—	—	44.5 · UD-Q3_K_XL	1
OpenLLaMA-3B	LLaMA	—	—	—	1

21 runs · last run Oct 4, 2026. Filter by family above; full board for every row.

VRAM tiers: <24 GB · 24–48 GB · >48 GB. A cell shows the best single-stream tok/s and quant recorded in that tier — not a guarantee for every loadout. Imported rows carry their original methodology link.

Related guides: best GPU for Qwen3 · quantization · Strix Halo 128GB