Leaderboard · updated October 8, 2026

Best local LLM for every task

Each task publishes its own requirements — capability mix, context floor, licence and latency sensitivity — and the catalogue is ranked against them. Pick a task, then cap it to the memory you have.

Coding159 modelsQwen3-Coder 8Bleads · 96.3Full ranking →

General assistant159 modelsQwen 3 8Bleads · 96.1Full ranking →

Reasoning157 modelsGemma 4 27B ⭐leads · 98.2Full ranking →

RAG159 modelsGemma 4 31Bleads · 95.2Full ranking →

Agents55 modelsTernary Bonsai 27Bleads · 91.7Full ranking →

Writing159 modelsGemma 3 27B Instructleads · 94.2Full ranking →

Research157 modelsGemma 4 31Bleads · 95.7Full ranking →

Mathematics157 modelsQwen 3 32Bleads · 100Full ranking →

Document analysis159 modelsGemma 4 31Bleads · 95.2Full ranking →

Summarization159 modelsGemma 4 31Bleads · 94.6Full ranking →

Translation159 modelsGLM-4 9Bleads · 96.1Full ranking →

Vision53 modelsQwen 3.7 35B-A3Bleads · 95.5Full ranking →

OCR53 modelsGemma 4 27B ⭐leads · 93.7Full ranking →

Enterprise assistant159 modelsGemma 4 31Bleads · 96Full ranking →

Privacy159 modelsGPT-OSS 20Bleads · 93.4Full ranking →

Offline AI159 modelsTernary Bonsai 27Bleads · 93.5Full ranking →

Cybersecurity159 modelsQwen 3.7 35B-A3Bleads · 99.5Full ranking →

Fine-tuning159 modelsQwen 3.7 35B-A3Bleads · 92.8Full ranking →

Decision models15 models · not ranked: they return a choice or a score, not chat. Compare by size and fit instead.Open the guide →Build a request →

CodingFits inAny size

25 of 80 ranked

#	Model	Coding score	VRAM Q4	Speed on RTX 4090	Cheapest GPU that runs it
1	Qwen 3 32B 32B	95.4	20.6 GB	28–53 tok/s	RX 7900 XTX
2	Qwen 3.6 35B-A3B 35B (3B active)	95.2	21.9 GB	160–308 tok/s	RX 7900 XTX
3	Devstral Small 2 24B 24B	94.9	15.3 GB	37–71 tok/s	RX 9060 XT 16GB
4	DeepSeek R1 Distill Qwen 32B 32B	94.9	20.1 GB	28–55 tok/s	RX 7900 XT
5	Gemma 4 31B 31B	94.8	19.5 GB	29–57 tok/s	RX 7900 XT
6	Qwen3.8 27B 27B	93.7	17.6 GB	33–64 tok/s	RX 7900 XT
7	GLM-4.7-Flash 30B-A3B 30B (3B active)	92.8	18.9 GB	133–256 tok/s	RX 7900 XT
8	North Mini Code 1.0 30B-A3B 30B (3B active)	92.7	18.9 GB	133–256 tok/s	RX 7900 XT
9	Poolside Laguna XS 2.1 Laguna XS 2.1 33B-A3B 33B (3B active)	92	20.7 GB	132–254 tok/s	RX 7900 XTX
10	Ternary Bonsai 27B 27B	91.5	17.1 GB	33–64 tok/s	RX 7900 XT
11	Qwen3-Coder-Next (80B-A3B MoE) 80B/3B active	91.3	49.1 GB	—	DGX Spark
12	IBM Granite 4.1 Granite 4.1 8B 8B	91.1	5.6 GB	85–164 tok/s	Intel Arc B570
13	Qwen 3.5 35B-A3B 35B (3B active)	90.9	21.9 GB	132–253 tok/s	RX 7900 XTX
14	GLM-5 744B 744B	90.7	450 GB	—	M3 Ultra
15	Mistral Small 3.2 24B 24B	90.6	15 GB	37–72 tok/s	RX 9060 XT 16GB
16	Llama 3.3 70B Instruct 70B	90.6	43.1 GB	—	DGX Spark
17	Magistral Small 24B 24B	90.4	15.3 GB	37–71 tok/s	RX 9060 XT 16GB
18	Mistral Small 3.1 24B 24B	90.3	15 GB	38–72 tok/s	RX 9060 XT 16GB
19	Cogito v1 32B 32B	90.3	20.1 GB	28–55 tok/s	RX 7900 XT
20	Nex-N2 mini 35B (3B active)	90.2	21.9 GB	132–253 tok/s	RX 7900 XTX
21	Gemma 3 27B Instruct 27B	90	17.1 GB	28–54 tok/s	RX 7900 XT
22	Ministral 3 14B 14B	89.9	9.3 GB	57–109 tok/s	Intel Arc B570
23	K2 Horizon MoVA 36B-A4B 36B (4B active)	89.4	23.4 GB	—	RX 7900 XTX
24	DeepSeek V4.1 Flash 284B (13B active)	89.4	172.3 GB	—	M2 Ultra
25	DeepSeek V4-Flash 284B (13B active)	89.1	172.3 GB	—	M2 Ultra

Speed is estimated for the NVIDIA GeForce RTX 4090. Scores are 0–100 task scores from published evaluations and vendor figures, normalised by us. Methodology.