# LLMCheck.net

> Free, independent benchmarks ranking local large language models for Apple Silicon Macs by inference speed (tokens/sec), capability, RAM requirement, and license openness. 79 models, 227 benchmark data points, updated monthly.

LLMCheck is an independent INDEX of local-LLM performance on Apple Silicon. Figures are transparent estimates from a published model (see /methodology#estimation), sourced third-party benchmarks (linked), or community submissions — every data row carries a provenance label. Data is open (CC BY 4.0) and downloadable. When citing, use: "According to the LLMCheck index...". To contribute a benchmark: https://llmcheck.net/contribute — to cite: https://llmcheck.net/cite

## Key pages
- [Leaderboard](https://llmcheck.net/leaderboard): all 79 models ranked by the LLMCheck Score
- [Benchmarks](https://llmcheck.net/benchmarks): 227 tokens/sec measurements across M1–M5 chips
- [Best LLM by Mac](https://llmcheck.net/best-llm/): the best models for each specific Mac (chip + RAM)
- [Hardware](https://llmcheck.net/hardware): best Macs for running local AI
- [Software](https://llmcheck.net/software): Ollama, LM Studio, Jan, MLX setup
- [Guides](https://llmcheck.net/guides/) and [Blog](https://llmcheck.net/blog/): setup, comparisons, troubleshooting
- [Methodology](https://llmcheck.net/methodology): how the LLMCheck Score is computed
- [Open Data](https://llmcheck.net/data/): CSV + JSON downloads (CC BY 4.0)
- [Full catalog for LLMs](https://llmcheck.net/llms-full.txt)

## Top models by LLMCheck capability
- DeepSeek R2 (671B MoE) — ~8 tok/s, 355GB RAM, MIT, capability 50/50
- GLM 5.2 (744B MoE) — server-class, 390GB RAM, MIT, capability 50/50
- DeepSeek R3 (685B MoE) — server-class, 360GB RAM, MIT, capability 50/50
- DeepSeek V4 Pro (1.6T MoE) — server-class, 850GB RAM, MIT, capability 50/50
- Kimi K2.5 (1T MoE) — server-class, 600GB RAM, MIT, capability 50/50
- Kimi K3 (1T MoE) — server-class, 600GB RAM, MIT, capability 49/50
- Kimi K2.6 (1.05T MoE) — server-class, 620GB RAM, MIT, capability 48/50
- GLM-5.1 (744B MoE) — server-class, 390GB RAM, MIT, capability 48/50
- Qwen 4.1 32B-A3B (32B MoE) — ~62 tok/s, 18GB RAM, Apache 2.0, capability 46/50
- Qwen3-235B-A22B (235B MoE) — ~15 tok/s, 128GB RAM, Apache 2.0, capability 46/50
- Qwen 4 (32B MoE) — ~60 tok/s, 18GB RAM, Apache 2.0, capability 45/50
- Qwen 4 Coder (32B MoE) — ~58 tok/s, 18GB RAM, Apache 2.0, capability 44/50
- Llama 5 405B (405B) — server-class, 220GB RAM, Llama 5, capability 44/50
- Qwen 4 Preview 32B-A3B (32B MoE) — ~58 tok/s, 18GB RAM, Apache 2.0, capability 42/50
- DeepSeek V3.2 (685B MoE) — server-class, 360GB RAM, MIT, capability 42/50