DATA-DRIVEN MODEL COMPARISON
Best local LLMs for reasoning
Compare science and competition-math results separately, with exact quantization sizes and source-linked test settings.
First Pareto choice on GPQA Diamond
Nanbeige 4.2 3B · 87.4 / 100 · 2.684 GB · Q4_K_M. This choice is calculated within this page’s eligible set.
GPQA Diamond: comparison
Hard science questions. This is a science reasoning test, not an overall intelligence score.
AIME 2026: comparison
Competition math from AIME 2026. Published answer accuracy.
| Model / quantization | Weights | Parameters | Score / 100 | Score / GB | Sources |
|---|---|---|---|---|---|
| Limite 1B - ViolettoBF16 Safetensors; custom code or vLLM plugin. Single-turn math specialist. | 2.071 GB | 1B | 94.01 | 45.398 | Get weights Score source |
| MiniCPM5 2B | 1.561 GB | 2.517B | 86.5 | 55.402 | Get weights Score source |
| Ternary Bonsai 2 27BCustom Prism llama.cpp kernels required; check backend support. | 5.947 GB | 26.896B | 95.83 | 16.115 | Get weights Score source |
| Spark X2.5 1.7B | 1.107 GB | 1.708B | 69.4 | 62.666 | Get weights Score source |
| K2 Horizon 0.9B | 0.666 GB | 1.078B | 48.5 | 72.803 | Get weights Score source |
| Swift 1.5 Qwen3.8-27B | 17.442 GB | 27.321B | 96 | 5.504 | Get weights Score source |
| Aleph Alpha Kolibri-1 | 47.454 GB | 78.103B | 96 | 2.023 | Get weights Score source |
| Qwen3.6-27B | 16.817 GB | 27.781B | 94.1 | 5.595 | Get weights Score source |
| Swift-Qwen3.8-27b | 17.442 GB | 27.781B | 94 | 5.389 | Get weights Score source |
| Qwen3.6-35B-A3B | 22.135 GB | 35.952B | 92.7 | 4.188 | Get weights Score source |
| Spark X2.5 4B | 2.6 GB | 4.112B | 90.7 | 34.882 | Get weights Score source |
| Xing4.0 29B-A4BPublished tests use very long contexts; this weight budget does not reproduce those settings. | 20.104 GB | 31.215B | 90 | 4.477 | Get weights Score source |
| Xing4.0-29B-A4B-GGUFPublished tests use very long contexts; this weight budget does not reproduce those settings. | 20.104 GB | 31.215B | 90 | 4.477 | Get weights Score source |
| gemma-4-31B-it | 18.324 GB | 31.273B | 89.2 | 4.868 | Get weights Score source |
| gemma-4-26B-A4B-it | 16.948 GB | 25.806B | 88.3 | 5.21 | Get weights Score source |
| Ternary Bonsai 27BCustom Prism llama.cpp kernels required; check backend support. | 7.165 GB | 26.896B | 87.5 | 12.212 | Get weights Score source |
| Bonsai 27B (1-bit)Custom Prism llama.cpp kernels required; check backend support. | 3.803 GB | 26.896B | 87.08 | 22.895 | Get weights Score source |
| gemma-4-12B-it | 7.122 GB | 11.96B | 77.5 | 10.882 | Get weights Score source |
| Gemma 4 E4B | 4.977 GB | 7.996B | 42.5 | 8.539 | Get weights Score source |
| Gemma 4 E2B | 3.107 GB | 5.123B | 37.5 | 12.071 | Get weights Score source |
How recommendations are calculated
We compare one benchmark at a time. A Pareto choice has no other scored model with both an equal or smaller weight file and an equal or higher score, with at least one strict improvement. The first choice is the frontier point nearest the chart’s top-left corner, using the same rounded memory scale and 0–100 score scale as the main chart. Other frontier choices follow, then other scored models by score.
Score per GB is shown for context; it does not choose the champion. Changing the budget or dataset can change the result. Scores are published point estimates, often for the source checkpoint rather than the selected quantization. They do not establish statistically significant differences or equal evaluation conditions.
RAM, VRAM and unified memory
Use one memory pool. System RAM is used for CPU inference; a dedicated GPU has separate VRAM. Do not add RAM and VRAM as if they were one GPU pool. Partial CPU offload needs a supported backend and can change speed. Unified memory is shared by CPU, GPU and the operating system; software can limit how much the GPU can allocate.
Our initial weight budget reserves 30% of system RAM or unified memory, with a minimum of 2 GB. For VRAM it reserves 20%, with a minimum of 1 GB. These are adjustable planning assumptions, not measured requirements. Context cache, runtime buffers, other programs, multimodal files and concurrent users can need more. A file that passes the budget check is a candidate, not a confirmed hardware-compatible configuration.
Check the model’s runtime and quantization support before downloading. See the Transformers cache documentation and llama.cpp backend documentation. GB uses decimal bytes; operating systems can report GiB instead.