Best Open-Weight Coding Models in 2026, Compared

There is no single leaderboard that ranks today's open-weight coding models fairly. Kimi K3.1 Pro, Mistral Large 4, DeepSeek V4.1 Flash and Qwen3.8 are each reported on different benchmark versions, inside different agent harnesses, mostly by the vendors themselves. Here is what each number actually says, which of these models you can download today, and how to pick one for your own codebase.

The short answer

• Highest independently measured agentic coding score: Kimi K3.1 Pro, at 67% on Terminal-Bench 4.0 in Artificial Analysis's launch run (Oct 23, 2026). Its weights are not downloadable yet; Moonshot says they will be by November 4.
• Cleanest license you can download now: DeepSeek V4.1 Flash, MIT-licensed on Hugging Face since October 17, 2026.
• Most practical to self-host: Qwen3.8-27B, a dense 27-billion-parameter model under Apache 2.0.
• Biggest recent jump: Mistral Large 4, which Mistral says moved from about 28% to 41% on Terminal-Bench 4.0 after three more weeks of reinforcement learning (reported Oct 29, 2026).

The scorecard, with its caveats attached

Every figure below names who reported it, when, and how. Read the rows as separate claims, not as one ranked table, because the setups differ.

• Kimi K3.1 Pro (Moonshot AI). Terminal-Bench 4.0: 65.7% reported by Moonshot with a Claude Code harness, and 67% measured by Artificial Analysis (Oct 23, 2026), the highest of any model it had tested. Artificial Analysis lists 2.8T total and 104B active parameters and a 1M-token context.
• Mistral Large 4 (Mistral AI). Terminal-Bench 4.0: 41%, per Mistral as reported by The Next Web (Oct 29, 2026), up from about 28% at the October 6 preview. Preview figures from Mistral: DeepSWE v1.1 61.7% and SWE-Atlas-QnA 59.4%. 1.05T total and 52B active parameters, 1M context, per Mistral's docs.
• DeepSeek V4.1 Flash (DeepSeek). Terminal-Bench 4.0: 42.3 with a Claude Code harness, DeepSWE v1.1: 77.0 with a Mini-SWE-V2 harness, NL2Repo: 68.2, all from DeepSeek's model card (Oct 17, 2026). About 763B parameters per the Hugging Face listing.
• Qwen3.8-2.4T-A95B (Alibaba). Terminal-Bench 2.1: 86.6 (Claude Code, average of 10 runs), SWE-bench Pro: 67.7, DeepSWE 1.1: 56.6, from Qwen's own table for its Max tier, released August 2026. 2.4T total and 95B active parameters.
• Qwen3.8-27B (Alibaba). Terminal-Bench 2.1: 73.0 with the Terminus harness, SWE-bench Pro: 61.7, DeepSWE 1.1: 42.2 (Claude Code harness), from Qwen's model card (August 2026).

These results are not directly comparable. Terminal-Bench 2.1 and Terminal-Bench 4.0 are different task sets, so Qwen's 86.6 and Kimi's 67 measure different things. Only three of the five models have a Terminal-Bench 4.0 number at all, and only Kimi's has been reproduced independently in our sources.

Why the numbers don't line up

Agentic coding benchmarks grade a model and its harness together: the loop that reads files, runs commands and decides when to stop. We covered how this works in SWE-bench vs Terminal-Bench, and this comparison shows the effect clearly.

• Same benchmark, different harness. DeepSeek reports DeepSWE v1.1 for V4.1 Flash with a Mini-SWE-V2 harness (77.0), while Qwen reports DeepSWE 1.1 for Qwen3.8-27B with Claude Code (42.2). Some of that 35-point gap may be the harness, not the model.
• Same model, different reporter. DeepSeek's own table puts V4-Pro-0813 at 87.9 on Terminal-Bench 2.1 using its own DeepSeek Harness at maximum effort. Independent runs of that model are not part of our sources, so treat the vendor number as an upper bound.
• Benchmark versions move fast. Kimi K3, the previous downloadable Kimi model, scores 12.6 on Terminal-Bench 4.0 in DeepSeek's comparison table. That is a reminder that last spring's open model and this autumn's are very different products.

For an independent cross-check, Artificial Analysis publishes a Coding Agent Index (version 2.0 combines FrontierCode v1.1, Terminal-Bench 4.0 and SWE-Atlas-QnA). Mistral says Large 4 now rivals GLM-5.3 on that index while using 47% fewer tokens per task; that is Mistral's description of a third-party score, as reported by The Next Web.

Kimi K3.1 Pro: top score, weights still pending

Kimi K3.1 Pro is the strongest agentic coder in this group on the evidence available. Artificial Analysis measured 67% on Terminal-Bench 4.0 and ranked it fourth on its overall Intelligence Index (Oct 23, 2026). We covered the launch in Kimi K3.1 Pro closes the open-weight gap.

The catch is in the word open. As of October 29, 2026, Moonshot had not published the weights or the license; it says weights will arrive by November 4. Until then, it is an API model with an open-weight promise. At about 2.8T total parameters, self-hosting it will be a multi-node job for most teams even after release.

Mistral Large 4: a big jump and fresh weights

Mistral Large 4 entered preview on October 6 with 28% on Terminal-Bench 4.0, well behind the leaders (see our Le Chonk coverage). On October 29, Mistral said extra reinforcement learning lifted that to 41%, just ahead of GLM-5.3's 40%, and added low and medium reasoning modes. It also says a quantized build runs on four B200 or B300 GPUs, and that smaller distilled versions are coming.

Mistral announced the open-weight release the same day. When we checked the Hugging Face repository that evening (Dubai time), it still showed as an upcoming release, so confirm the download and license terms before you plan a deployment around it.

DeepSeek V4.1 Flash: the easiest license

DeepSeek V4.1 Flash is the one model here that is both strong on vendor numbers and simple to adopt. It is MIT-licensed on Hugging Face, and DeepSeek's API now routes its deepseek-flash model name to it. DeepSeek reports 42.3 on Terminal-Bench 4.0 with a Claude Code harness, which, if it holds up independently, puts it level with Mistral Large 4 and well above Kimi K3.

At roughly 763B parameters it is not a laptop model. Its bigger sibling, DeepSeek-V4-Pro-0813 (1.6T total, 49B active, also MIT), reports higher scores, but on Terminal-Bench 2.1 rather than 4.0.

Qwen3.8: one giant, one you can actually run

Qwen3.8 comes in very different sizes. The 2.4T-parameter flagship posts Qwen's best coding numbers, but it uses the custom Qwen3.8-Max License: free for most uses, with a requirement to display the model name above 100 million monthly users or US$20 million monthly revenue, and a separate license for large model-as-a-service or AI coding assistant businesses with over US$50 million in revenue in 12 months.

Qwen3.8-27B is the practical pick. It is dense, multimodal, Apache 2.0, and reports 73.0 on Terminal-Bench 2.1 and 61.7 on SWE-bench Pro. As a rough guide, 27 billion parameters at 16-bit precision is about 54 GB of weights before the KV cache, which fits on one 80 GB data-center GPU; quantization brings it lower.

Licenses and self-hosting, as of Oct 29, 2026

• DeepSeek V4.1 Flash: MIT, weights on Hugging Face.
• Qwen3.8-27B: Apache 2.0, weights on Hugging Face.
• Qwen3.8-2.4T-A95B: Qwen3.8-Max License, permissive with scale-based conditions.
• Mistral Large 4: open-weight release announced Oct 29; repository still marked upcoming when we checked.
• Kimi K3.1 Pro: weights and license promised by Nov 4. Kimi K3 is downloadable today.

API prices

Prices as of Oct 29, 2026, per million tokens. Kimi K3.1 Pro costs $2 input, $10 output and $0.20 for cached input, as listed by Artificial Analysis; check the Kimi API docs for current rates. Mistral's Large 4 model page lists $1.36 input, $0.14 cached input and $4.18 output, though The Next Web reports final pricing has not been published yet. DeepSeek and Qwen publish rates on the DeepSeek API docs and Qwen Cloud. If you want to switch between these models without rewriting clients, an AI gateway helps; see our AI gateways comparison.

How to choose: run your own eval

Public scores shortlist models; your repository decides. A small private evaluation takes a day and answers the question the benchmarks can't.

1. Collect 20 to 50 recently closed issues or pull requests from your own repositories, with tests that prove the fix.
2. Pick one harness and keep it fixed. If you use Claude Code or a similar agent, point it at each model through an OpenAI- or Anthropic-compatible endpoint.
3. Run each model several times per task, because agent runs vary.
4. Track pass rate, cost per solved task and wall-clock time, not just pass rate.
5. Re-run when a model ships new weights. Mistral's 13-point jump in three weeks shows how quickly results go stale.

Verdict

If you need weights today with no license questions, start with DeepSeek V4.1 Flash, and use Qwen3.8-27B where a single GPU is the budget. Kimi K3.1 Pro has the best independently measured score and is worth testing through its API now, then again once the weights land. Mistral Large 4 has closed much of the gap and deserves a slot in your eval once its download is live. None of the cross-vendor numbers are apples to apples, so let your own tasks decide.

Sources

• DeepSeek: DeepSeek-V4.1-Flash-1017 model card (Oct 17, 2026)
• DeepSeek: DeepSeek-V4-Pro-0813 model card (Aug 13, 2026)
• Qwen: Qwen3.8-2.4T-A95B model card (Aug 2026)
• Qwen: Qwen3.8-Max License
• Qwen: Qwen3.8-27B model card (Aug 2026)
• Qwen: Qwen3.8 GitHub repository
• Moonshot AI: Kimi K3.1 Pro announcement
• Moonshot AI: Kimi K3.1 Pro quickstart
• Mistral AI: Mistral Large 4 announcement
• Mistral AI: Mistral Large 4 model docs
• Mistral AI: Mistral-Large-4-1T-A52B on Hugging Face
• DeepSeek API docs: first API call and models
• Artificial Analysis: Kimi K3.1 Pro launch benchmarks (Oct 23, 2026)
• Artificial Analysis: model leaderboards and Coding Agent Index
• The Next Web: Mistral Large 4 open weights release (Oct 29, 2026)