Best open-weight model for…
Pick the question you have. Each page below is computed from the benchmark results and says in one sentence how. A page names a best model only when at least 3 ranked models qualify; with fewer it says so. Scores come from a small fixed battery, so read the rank range as well as the place.
By memory
How much memory can you give a model? A model fits when its estimated memory need (the model file plus about 20% for context) is at most the budget. These use the same estimate and the same caveats as the hardware page.
- Best model for 8 GB: Gemma4 E4B IT, Overall 95.5 (143 ranked models fit). Typical machine: Entry laptop or 8 GB graphics card.
- Best model for 12 GB: Gemma 4 12B IT, Overall 98.6 (164 ranked models fit). Typical machine: 12 GB graphics card.
- Best model for 16 GB: Gemma 4 12B IT, Overall 98.6 (165 ranked models fit). Typical machine: 16 GB laptop or graphics card.
- Best model for 24 GB: gemma-4-31B-it, Overall 99.3 (202 ranked models fit). Typical machine: 24 GB graphics card.
- Best model for 32 GB: gemma-4-31B-it, Overall 99.3 (215 ranked models fit). Typical machine: 32 GB Mac or workstation.
- Best model for 48 GB: gemma-4-31B-it, Overall 99.3 (217 ranked models fit). Typical machine: 48 GB workstation card or Mac.
- Best model for 64 GB: gemma-4-31B-it, Overall 99.3 (219 ranked models fit). Typical machine: 64 GB Mac or workstation.
- Best model for 96 GB: gemma-4-31B-it, Overall 99.3 (222 ranked models fit). Typical machine: 96 GB unified-memory Mac.
- Best model for 128 GB: gemma-4-31B-it, Overall 99.3 (222 ranked models fit). Typical machine: 128 GB unified-memory Mac or mini PC.
By task and size
Each page shows the best models for one kind of task, for all sizes and then for each size class.
- Best model for coding: gpt-oss-120b, score 99.9 (227 models scored).
- Best model for reasoning: DeepSeek-R1-Distill-Qwen-32B, score 97.1 (227 models scored).
- Best model for math: Flint 40B A2B, score 100.0 (227 models scored).
- Best model for instruction following: Gemma 4 12B IT, score 100.0 (227 models scored).
- Best model for multilingual: Devstral Small 2 24B Instruct (2512), score 100.0 (227 models scored).
- Best model for tool use: Aya Expanse 8B, score 100.0 (227 models scored).
- Best model for writing: gemma-4-31B-it, score 7.0 (227 models scored).
- Best model for long context: DeepSeek-R1-Distill-Qwen-32B, score 100.0 (227 models scored).
- Best model for knowledge & honesty: Aya Expanse 32B, score 100.0 (227 models scored).
- Best model for safety: Apertus-8B-Instruct-2509, score 100.0 (227 models scored).
- Best model for pushback resistance: Cogito 8B V1 Preview Llama, score 100.0 (227 models scored).
- Best model for famous gotchas: gemma-4-26B-A4B-it, score 100.0 (227 models scored).
By machine
Pick the machine you have. Only the memory carries over to it: each page counts a share of a unified-memory machine's RAM as available to a model, says how much, and shows speeds as measured on the benchmark's own laptop, never as a speed for your machine.
- Best model for a Mac with 16 GB of memory: Gemma 4 12B IT, Overall 98.6 (147 ranked models fit in 9 GB).
- Best model for a Mac with 24 GB of memory: Gemma 4 12B IT, Overall 98.6 (165 ranked models fit in 14 GB).
- Best model for a Mac or Ryzen AI machine with 32 GB of memory: Gemma 4 12B IT, Overall 98.6 (175 ranked models fit in 19 GB).
- Best model for a Mac with 48 GB of memory: gemma-4-31B-it, Overall 99.3 (211 ranked models fit in 28 GB).
- Best model for a Mac or Ryzen AI machine with 64 GB of memory: gemma-4-31B-it, Overall 99.3 (217 ranked models fit in 38 GB).
- Best model for a Mac with 128 GB of memory: gemma-4-31B-it, Overall 99.3 (219 ranked models fit in 76 GB).
- Best model for an 8 GB graphics card: Gemma4 E4B IT, Overall 95.5 (143 ranked models fit in 8 GB).
- Best model for a 12 GB graphics card: Gemma 4 12B IT, Overall 98.6 (164 ranked models fit in 12 GB).
- Best model for a 16 GB graphics card: Gemma 4 12B IT, Overall 98.6 (165 ranked models fit in 16 GB).
- Best model for a 24 GB graphics card: gemma-4-31B-it, Overall 99.3 (202 ranked models fit in 24 GB).
- Best model for a Ryzen AI machine with 128 GB of memory: gemma-4-31B-it, Overall 99.3 (219 ranked models fit in 76 GB).
By task and memory
The best coding, tool use, multilingual and long-context models that fit in 8, 16 or 32 GB, with the same memory estimate as above.
- Best model for coding under 8 GB: Gemma4 E4B IT, score 89.9 (143 models scored).
- Best model for coding under 16 GB: Gemma 4 12B IT, score 98.2 (165 models scored).
- Best model for coding under 32 GB: Gemma 4 12B IT, score 98.2 (215 models scored).
- Best model for tool use under 8 GB: Aya Expanse 8B, score 100.0 (143 models scored).
- Best model for tool use under 16 GB: Aya Expanse 8B, score 100.0 (165 models scored).
- Best model for tool use under 32 GB: Aya Expanse 8B, score 100.0 (215 models scored).
- Best model for multilingual under 8 GB: gemma-2-9b-it, score 100.0 (143 models scored).
- Best model for multilingual under 16 GB: Gemma 4 12B IT, score 100.0 (165 models scored).
- Best model for multilingual under 32 GB: Devstral Small 2 24B Instruct (2512), score 100.0 (215 models scored).
- Best model for long context under 8 GB: Gemma4 E2B IT, score 100.0 (143 models scored).
- Best model for long context under 16 GB: Gemma4 E2B IT, score 100.0 (165 models scored).
- Best model for long context under 32 GB: DeepSeek-R1-Distill-Qwen-32B, score 100.0 (215 models scored).
Speed and value
- Fastest model that is still good: Qwen3-VL-30B-A3B-Instruct, 26.9 tokens per second (measured on Vulkan; speeds on different backends are not directly comparable) (54 models in the quality window).
- Best value for memory: Smollm2 135M Instruct, 159.66 Overall points per estimated GB.
Looking for something else? Every category has its own leaderboard, and you can compare any two models.