OPEN-WEIGHT MODELS · UPDATED NOVEMBER 2026
Best local LLMs for codingPicks by GPU size, with scores, VRAM and license
The best open-weight coding model is the strongest one your GPUs can hold. These are our picks for each size of hardware, with the coding 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
61.7% on SWE-bench Pro and 73.0% on Terminal-Bench 2.1 in about 24 GB of memory, under Apache 2.0. That is well ahead of Qwen3.6-27B (53.5% on SWE-bench Pro), which is the pick if you only have 24 GB.
Qwen3.8-27B hardware requirements →TWO GPUS · ABOUT 190 GB
MiMo-V2.6-Flash
Xiaomi · About 188 GB · MIT
87.6% on Terminal-Bench 2.1 and 67.9% on DeepSWE, level with GLM-5.3 (88.2% and 66.9%) in well under half the memory. MIT license.
ONE 8-GPU SERVER · ABOUT 537 GB
DeepSeek-V4.1-Flash-1017
DeepSeek · About 537 GB · MIT
77.0% on DeepSWE, the highest score of any open-weight model listed here, and 42.3% on the newer Terminal-Bench 4.0. MIT license. GLM-5.3 is the alternative if you have four H200 GPUs.
MORE THAN ONE SERVER · ABOUT 1.64 TB
Kimi K3.1 Pro
Moonshot AI · About 1.64 TB · Kimi K3.1 Pro License
65.7% on Terminal-Bench 4.0, far ahead of every other open-weight model listed, and 75.5% on DeepSWE. 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. Where a card lists several reasoning settings, the highest is shown. Developers use different test harnesses, so compare models within a column loosely. A dash means the developer has not published that figure.
| Model | SWE-bench Verified | SWE-bench Pro | Terminal-Bench | DeepSWE | Memory, 1 user | Smallest setup for 16 users | License |
|---|---|---|---|---|---|---|---|
| Qwen3.6-27B Qwen model card | 77.2% | 53.5% | 59.3% (v2.0) | — | About 20 GB | 1 NVIDIA L40S | Apache 2.0 |
| Qwen3.8-27B Qwen model card | — | 61.7% | 73.0% (v2.1) | 42.2% | About 24 GB | 1 NVIDIA L40S | Apache 2.0 |
| Qwen3.6-35B-A3B Qwen model card | 73.4% | 49.5% | 51.5% (v2.0) | — | About 24 GB | 1 NVIDIA L40S | Apache 2.0 |
| Nemotron 3.5 Super VL NVIDIA model card | 71.1% | 51.3% | 62.8% (v2.1) | — | About 80 GB | 1 NVIDIA RTX PRO 6000 Blackwell | OpenMDW 1.1 |
| MiniMax-M2.5 MiniMax model card | 80.2% | — | — | — | About 135 GB | 2 × NVIDIA H100 NVL | Modified MIT |
| DeepSeek-V4-Flash DeepSeek model card | 79.0% | 52.6% | 56.9% (v2.0) | — | About 177 GB | 2 × NVIDIA DGX Spark | MIT |
| MiMo-V2.6-Flash Xiaomi model card | — | — | 87.6% (v2.1) | 67.9% | About 188 GB | 2 × NVIDIA DGX Spark | MIT |
| GLM-5.2 Z.ai model card | — | 62.1% | 81.0% (v2.1) | 46.2% | About 455 GB | 4 × NVIDIA H200 NVL | MIT |
| GLM-5.3 Z.ai model card | — | — | 88.2% (v2.1) | 66.9% | About 460 GB | 4 × NVIDIA H200 NVL | GLM-5.3 License (MIT-like) |
| DeepSeek-V4.1-Flash-1017 DeepSeek model card | — | — | 42.3% (v4.0) | 77.0% | About 537 GB | 8 × NVIDIA H100 NVL | MIT |
| MiMo-V2.6-Pro Xiaomi model card | — | — | 89.9% (v2.1) | 71.9% | About 603 GB | 8 × NVIDIA H100 NVL | MIT |
| Kimi K2.6 Moonshot AI model card | 80.2% | 58.6% | 66.7% (v2.0) | — | About 626 GB | 8 × NVIDIA H100 NVL | Modified MIT |
| DeepSeek-V4-Pro DeepSeek model card | 80.6% | 55.4% | 67.9% (v2.0) | — | About 976 GB | 8 × NVIDIA H200 NVL | MIT |
| Kimi K3 Moonshot AI model card | — | — | 88.3% (v2.1) | 67.5% | About 1.62 TB | 8 × AMD Instinct MI325X | Kimi K3 License |
| Kimi K3.1 Pro Moonshot AI model card | — | — | 65.7% (v4.0) | 75.5% | 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
SWE-bench Verified. Fixing real GitHub issues in Python projects. The share of 500 human-checked issues the model resolves.
SWE-bench Pro. Harder, longer issues across more languages and larger codebases than SWE-bench Verified.
Terminal-Bench. Completing tasks in a real terminal: building, debugging and configuring software. Versions 2.0, 2.1 and 4.0 are not directly comparable, and 4.0 is much harder.
DeepSWE. A newer, harder software engineering benchmark where an agent may work on a task for hours. So far developers report it for their largest models.
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 coding: questions
What is the best local LLM for coding?
On a single GPU, Qwen3.8-27B: 61.7% on SWE-bench Pro in about 24 GB. On two GPUs, MiMo-V2.6-Flash. On one 8-GPU server, DeepSeek-V4.1-Flash-1017 at 77.0% on DeepSWE. With more than one server, Kimi K3.1 Pro.
What is the best local coding LLM for a 24 GB GPU?
Qwen3.6-27B at 4-bit needs about 20 GB for one user with 8K context, so it fits a 24 GB card with little room for long files. Qwen3.8-27B needs about 24 GB at NVFP4, so it is the better choice on a 32 GB or 48 GB card.
How much VRAM does a coding LLM need?
More than a chat model of the same size, because coding agents read whole files and long histories. Each developer needs their own KV cache, and it grows with context length. Plan for the context you will really use, not 8K.
Is a local coding model as good as a cloud one?
The largest open-weight models are close. On Terminal-Bench 2.1, Kimi K3 and GLM-5.3 score within a point of the cloud models their developers compare them with. Models that fit one GPU score lower, but your code never leaves your building.
Can a whole engineering team share one model?
Yes. One model serves many developers at once, with memory growing by each developer’s context. A model too large for one server has to be split across several, which Vanik Edge OS handles, along with keeping it running if a server fails.
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 harnesses, so small differences between models are not meaningful.
