OPEN-WEIGHT MODELS · UPDATED NOVEMBER 2026
Best local LLMs for reasoning and mathPicks by GPU size, with scores, VRAM and license
Reasoning models work through science, math and analysis problems step by step before they answer. These are our picks for each size of hardware, with the scores each developer publishes, the GPU memory each model needs and its license.
Our picks by hardware size
ONE GPU · 32 TO 94 GB
Qwen3.8-27B
Qwen · About 24 GB · Apache 2.0
89.2% on GPQA Diamond and 30.8% on Humanity’s Last Exam in about 24 GB of memory, ahead of Qwen3.6-27B (87.8% and 24.0%) and level with Nemotron 3.5 Super VL, which needs about 80 GB.
Qwen3.8-27B hardware requirements →TWO TO FOUR GPUS · ABOUT 177 GB
DeepSeek-V4-Flash
DeepSeek · About 177 GB · MIT
88.1% on GPQA Diamond, 34.8% on Humanity’s Last Exam and 88.4% on IMO-AnswerBench, under the MIT license with a 1M-token context.
ONE 8-GPU SERVER · ABOUT 455 GB
GLM-5.2
Z.ai · About 455 GB · MIT
99.2% on AIME 2026, 91.0% on IMO-AnswerBench and 40.5% on Humanity’s Last Exam, under the plain MIT license. DeepSeek-V4.1-Flash-1017 scores a little higher on Humanity’s Last Exam (42.6%) but publishes no math scores.
GLM-5.2 hardware requirements →MORE THAN ONE SERVER · ABOUT 1.64 TB
Kimi K3.1 Pro
Moonshot AI · About 1.64 TB · Kimi K3.1 Pro License
94.2% on GPQA Diamond and 51.1% on Humanity’s Last Exam, the highest open-weight scores listed here. Check its license before commercial use.
Choosing a model for your team?
Tell us your users, your work and your data. We will recommend a model and the hardware to run it inside your own building with Vanik Edge OS. Free, no obligation.
Scores and hardware, model by model
Scores are the ones each developer reports on its own model card, without tools. Where a card lists several reasoning settings, the highest is shown. A dash means the developer has not published that figure.
| Model | GPQA Diamond | Humanity’s Last Exam | AIME | IMO-AnswerBench | Memory, 1 user | Smallest setup for 16 users | License |
|---|---|---|---|---|---|---|---|
| Qwen3.6-27B Qwen model card | 87.8% | 24.0% | 94.1% (2026) | 80.8% | About 20 GB | 1 NVIDIA L40S | Apache 2.0 |
| Gemma 4 31B Google model card | 84.3% | 19.5% | 89.2% (2026) | — | About 20 GB | 1 NVIDIA L40S | Apache 2.0 |
| Qwen3.8-27B Qwen model card | 89.2% | 30.8% | — | — | About 24 GB | 1 NVIDIA L40S | Apache 2.0 |
| Qwen3.6-35B-A3B Qwen model card | 86.0% | 21.4% | 92.7% (2026) | 78.9% | About 24 GB | 1 NVIDIA L40S | Apache 2.0 |
| Nemotron 3.5 Super VL NVIDIA model card | 89.9% | 31.4% | — | — | About 80 GB | 1 NVIDIA RTX PRO 6000 Blackwell | OpenMDW 1.1 |
| MiniMax-M2.5 MiniMax model card | 85.2% | 19.4% | 86.3% (2025) | — | About 135 GB | 2 × NVIDIA H100 NVL | Modified MIT |
| DeepSeek-V4-Flash DeepSeek model card | 88.1% | 34.8% | — | 88.4% | About 177 GB | 2 × NVIDIA DGX Spark | MIT |
| GLM-5.2 Z.ai model card | 91.2% | 40.5% | 99.2% (2026) | 91.0% | About 455 GB | 4 × NVIDIA H200 NVL | MIT |
| DeepSeek-V4.1-Flash-1017 DeepSeek model card | — | 42.6% | — | — | About 537 GB | 8 × NVIDIA H100 NVL | MIT |
| Kimi K2.6 Moonshot AI model card | 90.5% | 34.7% | 96.4% (2026) | 86.0% | About 626 GB | 8 × NVIDIA H100 NVL | Modified MIT |
| DeepSeek-V4-Pro DeepSeek model card | 90.1% | 37.7% | — | 89.8% | About 976 GB | 8 × NVIDIA H200 NVL | MIT |
| Kimi K3 Moonshot AI model card | 93.5% | 43.5% | — | — | About 1.62 TB | 8 × AMD Instinct MI325X | Kimi K3 License |
| Kimi K3.1 Pro Moonshot AI model card | 94.2% | 51.1% | — | — | About 1.64 TB | 8 × AMD Instinct MI325X | Kimi K3.1 Pro License |
Most of these models need several GPUs working as one. Buying the GPUs is the easy part. Vanik Edge OS splits the model across them, keeps it running if a server fails and manages it all from one console. See how Edge OS clusters GPU servers
What the benchmarks measure
GPQA Diamond. Graduate-level multiple-choice questions in physics, chemistry and biology.
Humanity’s Last Exam. Expert-written questions across many fields, built to be at the edge of what models can answer. Scores are low for every model.
AIME. Problems from the American Invitational Mathematics Examination. The 2025 and 2026 sets are different papers.
IMO-AnswerBench. Olympiad-level math problems with short answers that can be checked.
FROM SIZING TO RUNNING
Sizing the hardware is the easy part.
Getting several GPUs to work as one system for a whole company is where most projects slow down. That is the part Vanik Edge OS does for you.
STEP 01 · THIS TOOL
Size it
Work out the GPU memory your models need and which GPUs or servers can hold them.
STEP 02 · VANIK EDGE OS
Cluster it
Edge OS joins your GPU servers into one cluster, splits a large model across them and keeps it running if a server fails. NVIDIA and AMD servers work side by side.
STEP 03 · VANIK EDGE OS
Run it
Your team gets AI apps with single sign-on and an audit trail, managed from one console, with your data staying inside your building.
Best local LLMs for reasoning and math: questions
What is the best local LLM for reasoning?
On a single GPU, Qwen3.8-27B. On two to four GPUs, DeepSeek-V4-Flash. On one 8-GPU server, GLM-5.2. With more than one server, Kimi K3.1 Pro, at 94.2% on GPQA Diamond.
What is the best local LLM for math?
GLM-5.2 has the highest published math scores here: 99.2% on AIME 2026 and 91.0% on IMO-AnswerBench. On a single GPU, Qwen3.6-27B scores 94.1% on AIME 2026.
Do reasoning models need more GPU memory?
Not for the weights, but they write long chains of thinking tokens before the answer, which fills the context and takes longer. Plan for a larger context per user and slower replies at the highest reasoning setting.
Why is gpt-oss-120b not in the table?
Its model card on Hugging Face does not publish benchmark figures, and this table only uses figures from each model’s own card. It is still a sound choice for text reasoning on one 80 GB-class GPU.
Can I turn reasoning off to save time?
On most of these models, yes. DeepSeek-V4-Pro, for example, scores 90.1% on GPQA Diamond at its highest setting and 72.9% with thinking off. Use the lower setting for simple questions and the higher one for hard problems.
Where do these scores come from?
Each developer’s own model card, linked in the table. We did not run the benchmarks ourselves, and developers use different test settings, so small differences between models are not meaningful.
