TokenAssemble

Browse · 40 tracked

Local LLMs

We track 40 open-weight models from 7 developers — 7 MoE among them — with measured GGUF sizes per quant, each graded against all 50 machines we cover. “Runs on” counts machines where the model loads cleanly (grade B or better) at a quality quant with llama.cpp at 8K context — see how the math works.

ModelParamsArchitectureMax contextQuantsRuns on
DeepSeek-R1671BMoE · 37B active160K4none
DeepSeek-V4-Flash284BMoE · 13B active1024K42/50
gpt-oss-120b117BMoE · 5.1B active128K16/50
Qwen2.5-72B-Instruct72.7Bdense32K47/50
Llama-3.1-70B-Instruct70.6Bdense128K47/50
Llama-3.3-70B-Instruct70.6Bdense128K47/50
Qwen3.6-35B-A3B35BMoE · 3B active256K416/50
Qwen2.5-32B-Instruct32.8Bdense32K416/50
QwQ-32B32.8Bdense40K416/50
Qwen3-32B32.8Bdense40K416/50
DeepSeek-R1-Distill-Qwen-32B32.8Bdense128K416/50
Qwen3-30B-A3B30.5BMoE · 3.3B active40K420/50
Qwen3.6-27B27.8Bdense256K420/50
gemma-3-27b-it27.4Bdense128K416/50
gemma-2-27b-it27.2Bdense8K420/50
gemma-4-26B-A4B-it26.5BMoE · 3.8B active256K420/50
Mistral-Small-24B-Instruct-250123.6Bdense32K421/50
gpt-oss-20b21BMoE · 3.6B active128K136/50
DeepSeek-R1-Distill-Qwen-14B14.8Bdense128K436/50
Qwen2.5-14B-Instruct14.8Bdense32K436/50
Qwen3-14B14.8Bdense40K436/50
Phi-414.7Bdense16K436/50
Mistral-Nemo-Instruct-240712.3Bdense128K445/50
gemma-3-12b-it12.2Bdense128K436/50
Qwen3.5-9B9.7Bdense256K446/50
gemma-2-9b-it9.2Bdense8K446/50
Qwen3-8B8.2Bdense40K450/50
DeepSeek-R1-Distill-Llama-8B8Bdense128K450/50
Llama-3.1-8B-Instruct8Bdense128K450/50
DeepSeek-R1-Distill-Qwen-7B7.6Bdense128K450/50
Qwen2.5-7B-Instruct7.6Bdense32K450/50
Mistral-7B-Instruct-v0.37.3Bdense32K450/50
gemma-3-4b-it4.3Bdense128K450/50
Phi-3.5-mini-instruct3.8Bdense128K450/50
Llama-3.2-3B-Instruct3.2Bdense128K450/50
Qwen2.5-3B-Instruct3.1Bdense32K450/50
gemma-2-2b-it2.6Bdense8K450/50
Qwen2.5-1.5B-Instruct1.5Bdense32K450/50
Llama-3.2-1B-Instruct1.2Bdense128K450/50
Qwen2.5-0.5B-Instruct0.5Bdense32K450/50

Params, layers, and context come from each repo's config.json; quant sizes are measured from the published GGUF files. “none” is an honest answer, not a data gap — some models simply don't fit consumer hardware.