The best open source LLM for coding right now is Kimi K3.1 Pro, from Moonshot AI: 67.2% on Artificial Analysis's Terminal-Bench 4.0, read on 20 November 2026, the top score on that chart, open or hosted.

Need a plain MIT licence? Take DeepSeek V4.1 Flash 1017. It is the best open model on FrontierCode-AA 2.0 at 28.5%, the board that grades whether a maintainer would merge the change, and second among open models on Arena's WebDev board.

On this page
  1. The short answer
  2. How I picked them
  3. Quick comparison
  4. 1. Kimi K3.1 Pro
  5. 2. DeepSeek V4.1 Flash 1017
  6. 3. GLM-5.3
  7. 4. MiMo-V2.6-Pro
  8. 5. Inkling 2 Small
  9. 6. Qwen3.8-27B
  10. 7. Gemma 4
  11. Which one should you run?
  12. Open vs hosted
  13. FAQ
  14. Sources

I read each model's Hugging Face card and licence file, the labs' release posts, and four coding leaderboards (Artificial Analysis, Arena, Terminal-Bench and SWE-bench) on 20 November 2026. No model was downloaded or run for this post. Leaderboards move, so every score below carries its board and its date, and no number from one board is set against a number from another.

A word on the label. All seven are open-weight models: five put their weights under an OSI-approved licence, and two use their makers' own.

TL;DR

The short answer

Every score read 20 November 2026, on the board named:

  1. Best overall: Kimi K3.1 Pro scores 67.2% on Artificial Analysis's Terminal-Bench 4.0, the top score on that chart, open or hosted. Its licence is Moonshot's own, with conditions for big model-hosting businesses.
  2. Best open source coding model with an MIT licence: DeepSeek V4.1 Flash 1017, the best open model on FrontierCode-AA 2.0 (28.5%) and second of 63 open models on Arena's WebDev board (1,659).
  3. The small Apache 2.0 picks: Gemma 4, which Google says runs quantized on consumer GPUs, and Qwen3.8-27B, a 27B dense model that is an 18GB download in Ollama. Small models score far lower on agentic tests: Qwen3.8-27B gets 5.6% on the same Terminal-Bench 4.0.
  4. Hosted models still lead on three of the four current boards: 28.5% against 49.1% on FrontierCode-AA 2.0, and Kimi K3.1 Pro ranks 17th on Arena WebDev and 10th on Terminal-Bench 4.1.
  5. Read the licence: GLM-5.1 and GLM-5.2 are MIT, GLM-5.3 uses Z.ai's own licence, and Mistral went from Apache 2.0 on Large 3 to its own licence on Large 4.

How I picked them

Four current boards did the ranking, each with a different job. I read all four on 20 November 2026:

  • Artificial Analysis's Terminal-Bench 4.0 is the main board. AA's method line: "We run all 66 Terminal-Bench 4.0 tasks with the mini-swe-agent harness and report pass@1 averaged over three repeats per task." Same tasks and the same agent for every model, so its scores compare.
  • FrontierCode-AA 2.0 checks merge quality. It is AA's version of Cognition's FrontierCode 2.0: 132 tasks in real open-source repositories, "graded on whether a maintainer would merge the agent's change." AA runs every task three times with its Stirrup agent harness.
  • Arena's Code Arena WebDev adds human votes. It holds 1,036,937 votes on 163 models, and its licence filter leaves 63 open source models.
  • Terminal-Bench 4.1 on tbench.ai shows where the hosted tools sit. Each row names the agent it ran in, so its rows are not like for like.
Artificial Analysis's Terminal-Bench 4.0 page captured on 20 November 2026, showing a summary card above a bar chart of 28 models in which Kimi K3.1 Pro comes first at 67.2%.
Artificial Analysis's Terminal-Bench 4.0 chart, read on 20 November 2026. It runs every model with the same mini-swe-agent harness; open-weight Kimi K3.1 Pro tops it at 67.2%.

SWE-bench Verified has no entries after 26 February 2026, so I did not rank by it. Aider's polyglot board (newest run October 2025) and LiveCodeBench's data (last modified August 2025) are older still.

The entry rule: downloadable weights on Hugging Face and a licence file I could read. The Open Source Initiative's Open Source AI Definition 1.0 asks for "Data Information", "Code" and "Parameters"; I checked licences only, so I call all seven open-weight and note which licences sit on the OSI's approved list.

Read, but left out:

  • Mistral Large 4: its licence changed from Large 3's Apache 2.0 to Mistral's own.
  • Ling 3.1 Flash 1019: I did not read its licence.
  • Qwen3.8-Flash-Next: Qwen Community License 1.0.
  • Tencent's Hy4 preview: Arena's open filter lists it, but Artificial Analysis marks it as not open weights.
  • gpt-oss: on neither Arena's open list nor AA's charts.
  • Devstral 2 and Granite 4.1 8B: the bottom two of Arena's 63 open models.

Quick comparison

Seven open-weight coding models, from their Hugging Face cards, licence files and four leaderboards, read 20 November 2026.
ModelMakerLicenceSize (total / active)Best coding signal (board)What you need (lab's or runner's words)
Kimi K3.1 ProMoonshot AICustom: Kimi K3.1-Pro License ($20M Model-as-a-Service clause, display clause)2.8T / 104BAA Terminal-Bench 4.0: 67.2%, top of the chart1,561 GB of weights; server engines (vLLM, SGLang, TokenSpeed)
DeepSeek V4.1 Flash 1017DeepSeekMIT (OSI-approved)552B backbone (+196B memory) / 16BFrontierCode-AA 2.0: 28.5%, best open510 GB of weights
GLM-5.3Z.aiCustom: GLM-5.3 License ($10B Model-as-a-Service security review)753B / 40BAA Terminal-Bench 4.0: 41.9%756 GB of weights; text input only
MiMo-V2.6-ProXiaomiMIT (OSI-approved)1.02T / 42BAA Terminal-Bench 4.0: 34.8%Card examples: 8 GPUs (vLLM) or 2 nodes (SGLang)
Inkling 2 SmallThinking MachinesApache 2.0 (OSI-approved)594B / 17BArena WebDev open: 1,644 (#4)Card: at least 500GB aggregated VRAM (2x B300 or 4x H200)
Qwen3.8-27BAlibaba QwenApache 2.0 (OSI-approved)27B denseArena WebDev open: 1,589 (#13)Ollama download 18GB (bf16 56GB); no hardware line on the card
Gemma 4 31B / 26B A4BGoogle DeepMindApache 2.0 (OSI-approved)30.7B dense / 25.2B with 3.8B activeArena WebDev open: 1,370 / 1,358Google: bf16 fits one 80GB H100, quantized runs on consumer GPUs; Ollama 19GB / 16GB

Scores come from different boards, so compare a number only with others from the same board. Every cell read 20 November 2026; nothing here was tested.

1. Kimi K3.1 Pro: the strongest open model this month

Kimi K3.1 Pro is best for long, autonomous coding work, if you have a GPU cluster or pay a provider to run it. Moonshot's card calls it "an open-weight, natively multimodal agentic model, and our most capable 3T-class model to date." It was announced in late October 2026, with weights on Hugging Face from 2 November.

Kimi K3.1 Pro's Hugging Face model page captured on 20 November 2026, showing the kimi-k3.1-pro licence tag, a 2.8T parameter count and the Inference Providers panel.
Kimi K3.1 Pro's Hugging Face page on 20 November 2026: 2.8T parameters under Moonshot's own Kimi K3.1-Pro License.

It has 2.8T total parameters, 104B active, a 1,048,576-token context and 1,561 GB of weight files, already quantized (MXFP4 weights, MXFP8 activations); the card names vLLM, SGLang and TokenSpeed but gives no GPU count. Hugging Face lists seven inference providers, and Moonshot's API serves it as kimi-k3.1-pro.

Among open models it comes first on three of the four boards: first on Artificial Analysis's Terminal-Bench 4.0 chart at 67.2%, first of 63 open models on Arena's WebDev board at 1,683 (17th overall), and the best open-weight row on tbench.ai's Terminal-Bench 4.1, 10th overall at 59.7% inside Claude Code. On the first, Artificial Analysis says it "scores the highest on Terminal-Bench 4.0 with a score of 67.2%", ahead of Gemini 4 Carbon (High) at 65.2% and GPT-6.2 Astra (Max) at 64.1%. It slips on FrontierCode-AA 2.0: 25.5%, three points behind DeepSeek.

Arena's Code Arena WebDev leaderboard captured on 20 November 2026 with the licence filter set to Open Source, showing 63 open source models in a bar chart led by Kimi K3.1 Pro at 1,683.
Arena's Code Arena WebDev leaderboard filtered to open-source licences, read on 20 November 2026: Kimi K3.1 Pro leads at 1,683, ahead of DeepSeek V4.1 Flash at 1,659.

Kimi's own post reports 65.7% on Terminal-Bench 4.0 with "the Claude-Code harness". That is Kimi's number; I rank by Artificial Analysis's 67.2%.

Kimi's licence adds three clauses to an MIT-style grant.

  • A "Model as a Service" business with revenue over $20 million in any 12 consecutive months must sign a separate agreement with Moonshot AI before any commercial use.
  • A commercial product with "more than 100 million monthly active users", or more than $20 million in monthly revenue, must show "Kimi K3.1-Pro" prominently in its interface.
  • Neither rule applies to internal use, or to use through Moonshot's official products and certified partners.

2. DeepSeek V4.1 Flash 1017: best if you need an MIT licence

DeepSeek V4.1 Flash 1017 is the pick when the licence has to be MIT and the model still has to rank near the top. Its card says so in one line: "This repository and the model weights are licensed under the MIT License."

DeepSeek V4.1 Flash 1017's Hugging Face model page captured on 20 November 2026, showing the MIT licence tag, a 763B parameter widget and the card's opening sentence.
DeepSeek V4.1 Flash 1017 on Hugging Face, 20 November 2026: MIT-licensed weights. The 763B figure counts every tensor; DeepSeek's card gives 552B backbone parameters.

It wins the merge test. On FrontierCode-AA 2.0 it scores 28.5%, the best of any open model, and Arena's WebDev voters rank it second among open models at 1,659, 24 points behind Kimi. DeepSeek calls the 1017 release "a minor update to DeepSeek-V4.1-Flash"; it went up on Hugging Face on 17 October 2026, and Arena lists it as DeepSeek-V4.1-Flash-20261017.

Its size gets quoted two ways. DeepSeek's base card describes "a multimodal Mixture-of-Experts (MoE) model with 552B backbone parameters", 16B of them active, plus an add-on "Engram conditional memory" of 196B parameters. Hugging Face's widget counts every tensor and shows 763B. Artificial Analysis uses 552B, and so do I. Context runs to one million tokens.

DeepSeek's trade comes down to licence against size.

Pros

  • MIT licence on the weights
  • The best open score on FrontierCode-AA 2.0 (28.5%)

Cons

  • 510 GB of weights in 48 files, and no GPU line on the card
  • 33.8% on AA's Terminal-Bench 4.0, about half of Kimi K3.1 Pro's 67.2%

DeepSeek's card reports 42.3 on Terminal-Bench 4.0 and notes "We use the Claude Code harness for Terminal-Bench 4.0". Artificial Analysis's own run gives 33.8%. For local use, the card points to an inference folder for converting the weights.

3. GLM-5.3: second on both Terminal-Bench boards, under a new licence

GLM-5.3 from Z.ai is the agentic runner-up: 41.9% on Artificial Analysis's Terminal-Bench 4.0 and 41.2% on tbench.ai's 4.1 inside Claude Code, second among open models on both. Its 755.6 GB of weights come to under half of Kimi's download.

GLM-5.3's Hugging Face model page captured on 20 November 2026, showing the glm-5.3 licence tag, a 753B parameter count and Z.ai's coding claim at the top of the card.
GLM-5.3 on Hugging Face, 20 November 2026. Unlike GLM-5.1 and GLM-5.2, which are MIT, it ships under Z.ai's own GLM-5.3 License.

Read the licence before the scores. GLM-5.1 and GLM-5.2 are tagged MIT on Hugging Face, and GLM-5.3 ships under Z.ai's own GLM-5.3 License, though its card says it shares GLM-5.2's base model. The grant is MIT-style with one addition: a Model as a Service business whose revenue, with its affiliates, tops $10 billion over any 12 consecutive months "must pass Z.AI's security review before using the Software or its derivative works for any commercial purpose." Pages that still sell GLM as the MIT pick are describing GLM-5.1 or GLM-5.2.

One limit sits in Z.ai's docs: "GLM-5.3 currently supports text-only inputs, with a 1M-token context window and a maximum output length of 128K tokens." So no pasting a screenshot of the broken page. Z.ai's card calls it "the most capable open-weights model for coding", measured on Z.ai's in-house benchmark; both Terminal-Bench boards put Kimi ahead.

GLM-5.3-Flash keeps the MIT licence.

It has 320B total and 18B active parameters and a 328.3 GB download. It scores 37.0% on Terminal-Bench 4.1 inside Claude Code and 1,610 on Arena's open board, 11th, and the engines on its card include KTransformers and Unsloth.

4. MiMo-V2.6-Pro: Xiaomi's MIT model for big servers

MiMo-V2.6-Pro is the largest MIT-licensed model here and is built for a GPU cluster: 1.02T total parameters, 42B active, a 1M-token context and a 573.5 GB download. It scores 34.8% on Artificial Analysis's Terminal-Bench 4.0, third among open models on that chart, and 24.0% on FrontierCode-AA 2.0. The card's own serving examples split it across 8 GPUs in vLLM or two nodes in SGLang. Xiaomi publishes -RL and -MOPD checkpoints, and the boards do not say which one they scored.

The smaller MiMo-V2.6-Flash (309B total, 15B active, also MIT) rates 1,638 on Arena's open WebDev board, fifth, where Arena labels it the "Daily driver".

5. Inkling 2 Small: Apache 2.0, strong on web apps, weak in the terminal

Inkling 2 Small suits web-app generation and reviewable changes when you want Apache 2.0 on a large model. Small is relative. Thinking Machines' post calls it "a Mixture-of-Experts model with 594B total parameters, 17B active", and the card gives it a context window "of up to 1M tokens".

The weights came a week after the launch. The 5 November 2026 post said it "will initially be available through Tinker, with closed weights", and an update followed: "Nov 12, 2026 update: Inkling 2 Small’s weights are now available for download on HuggingFace."

Scores
Fourth among open models on Arena's WebDev board (1,644) and 25.4% on FrontierCode-AA 2.0, but 13.1% on Artificial Analysis's Terminal-Bench 4.0 and 13.0% on tbench.ai's 4.1 inside mini-swe-agent.
Hardware (card)
"The main NVFP4 checkpoint requires a GPU cluster with at least 500GB of aggregated VRAM", for example 2x NVIDIA B300 or 4x NVIDIA H200. The MXFP8 checkpoint needs "at least 800GB of aggregated VRAM".
Download
393.5 GB for the main NVFP4 repository.
Hosted
NVIDIA, Fireworks, Together, Baseten, Nebius, Modal and Cloudflare serve it, per the release post.

6. Qwen3.8-27B: a 27B Apache 2.0 model that out-rates Gemma 4 on Arena's open board

Qwen3.8-27B is the best-rated small dense model I checked: 1,589 on Arena's open WebDev board, 13th of 63, above both Gemma 4 sizes. Its card calls it "a compact, deployment-friendly dense model" and "a native vision-language model", with "Context Length: 262,144 natively and extensible up to 1,000,000 tokens."

Qwen3.8-27B's Hugging Face model page captured on 20 November 2026, showing the apache-2.0 licence tag and a 28B parameter widget.
Qwen3.8-27B on Hugging Face, 20 November 2026: a 27B dense model under Apache 2.0.

Ollama's library page lists these downloads:

  • qwen3.8:27b, the default tag: 18GB, with 256K context
  • qwen3.8:27b-mxfp8: 30GB
  • qwen3.8:27b-bf16: 56GB
  • 27b-mlx, the MLX build: 18GB

Then the number none of the ranking pages I opened prints. On Artificial Analysis's Terminal-Bench 4.0 it scores 5.6%, against 67.2% for Kimi K3.1 Pro, and 14.9% on FrontierCode-AA 2.0. By Arena's votes it is good for chat-style help and generating web apps. For long autonomous terminal tasks it is weak.

Only the 27B is Apache 2.0 among the Qwen3.8 models I read.

Qwen3.8-Flash-Next uses the Qwen Community License 1.0, under which a "Model as a Service or AI Work Assistant business" must "obtain a separate license from Qwen", and very large products must display the model name. The Qwen3.8-Max open weights carry their own qwen3.8-max licence. Check the tag on the exact repository before you download.

7. Gemma 4: the smallest footprint, from Google, under Apache 2.0

Gemma 4 is the pick for a single machine, and the only model here whose maker says it fits one. Google's launch post: "our unquantized bfloat16 weights fit efficiently on a single 80GB NVIDIA H100 GPU. For local setups, quantized versions run natively on consumer GPUs to power your IDEs, coding assistants and agentic workflows."

Gemma 4 31B's Hugging Face model page captured on 20 November 2026, showing the Apache 2.0 licence and a table of the five Gemma 4 sizes.
Gemma 4 31B on Hugging Face, 20 November 2026. Google lists five sizes under Apache 2.0, from E2B to 31B.

Two sizes matter for coding, the 31B dense and the 26B A4B Mixture-of-Experts, both with 256K tokens of context. Ollama's qat tags (gemma4:31b-it-qat, gemma4:26b-a4b-it-qat) download at 19GB and 16GB, against 63GB and 52GB in bf16. Google also ships E2B and E4B for mobile and edge devices, and a 12B.

The licence needs no footnote. Google: "That's why Gemma 4 is released under a commercially permissive Apache 2.0 license."

The scores are the weak spot. Arena's open WebDev board puts the 31B 41st at 1,370 and the 26B A4B 45th at 1,358, and its open view labels the 31B "Efficient". Gemma 4 is not on Artificial Analysis's Terminal-Bench 4.0 or FrontierCode-AA 2.0 charts, so there is no agentic score to quote. Google's own model card reports 80.0% on LiveCodeBench v6 for the 31B, a vendor figure on a benchmark whose public results file was last modified in August 2025. Released on 2 April 2026, it is the oldest pick on this list.

Which one should you run? The best local LLM for coding depends on your hardware and licence

Start from the machine you have. Then check the licence against what you plan to ship.

For one consumer GPU, Gemma 4 is the only pick whose maker says it fits.

Google says quantized Gemma 4 runs natively on consumer GPUs. Qwen3.8-27B rates higher on Arena's open board but has no hardware line on its card, and neither gives a VRAM figure for a consumer GPU, so I give none.

With one 80GB data-centre GPU, Gemma 4 31B fits without quantization.

Google says the unquantized bfloat16 weights fit on a single 80GB NVIDIA H100. No other maker among the seven makes a single-GPU claim.

The five big models need a multi-GPU server or someone else's.

Their weight files alone, as listed on Hugging Face:

  • Kimi K3.1 Pro: 1,561 GB
  • GLM-5.3: 756 GB
  • MiMo-V2.6-Pro: 573.5 GB
  • DeepSeek V4.1 Flash 1017: 510 GB
  • Inkling 2 Small: 393.5 GB, and a card that asks for a GPU cluster with at least 500GB of aggregated VRAM

For most developers I would start with a hosted endpoint: Inkling's post names seven API hosts, and Hugging Face shows inference providers for Kimi K3.1 Pro, GLM-5.3 and Qwen3.8-27B. My page on free LLM API tiers tracks which providers give you a free allowance. If you still want your own box, read what self-hosting really takes first.

A download size is not a memory requirement. The GB figures in this post are summed weight files on Hugging Face or the runner's download size on Ollama. Only Thinking Machines' Inkling card and Google's Gemma 4 post state hardware, so read every other number as disk space.

If you will sell access to the model, read two licences first.

MIT and Apache 2.0 both sit on the OSI's approved-licence list. By licence alone, the seven split three ways:

  • MIT: DeepSeek V4.1 Flash 1017 and MiMo-V2.6-Pro, plus GLM-5.3-Flash.
  • Apache 2.0: Inkling 2 Small, Qwen3.8-27B and Gemma 4.
  • The maker's own licence: Kimi K3.1 Pro, with a separate agreement for model-hosting businesses over $20 million in 12 months and a name-display rule for very large products, and GLM-5.3, with a security review for hosts over $10 billion.

Both custom conditions apply before commercial use starts. Kimi's says the business "must enter into a separate agreement with Moonshot AI before using the Software or its derivative works for any commercial purpose." This is a summary and not legal advice; read the licence file in the repository you actually download.

A model still needs an agent around it before it writes code for you.

A model answers prompts; an agent reads your repository, runs commands and edits files. tbench.ai runs open models inside Claude Code and mini-swe-agent, and Kimi's card offers "OpenAI/Anthropic-compatible API" access. For the code around the model, my guide to open-source AI agent frameworks covers LangGraph, CrewAI and smolagents, and the framework leaderboard ranks 43 framework repositories.

Open models vs Claude Code, Codex and Copilot: the gap depends on the board

Ask whether open models have caught up and the answer changes with the board. Here is the best open model against the best hosted one on each, all read on 20 November 2026:

  • Artificial Analysis, Terminal-Bench 4.0: open leads. Kimi K3.1 Pro scores 67.2%, ahead of Gemini 4 Carbon (High) at 65.2%, with every model in the same mini-swe-agent harness.
  • FrontierCode-AA 2.0: open trails by about 20 points, 28.5% for DeepSeek V4.1 Flash 1017 against 49.1% for Claude Fable 5.2 (High, Default Fallback).
  • Arena Code Arena WebDev, all models: Kimi K3.1 Pro is 17th at 1,683, and Claude Fable 5.2 (xHigh) leads at 1,823.
  • Terminal-Bench 4.1 on tbench.ai: Kimi K3.1 Pro is 10th at 59.7% inside Claude Code; Copilot Fusion (max) leads at 71.2% inside GitHub Copilot.
The Terminal-Bench 4.1 leaderboard on tbench.ai captured on 20 November 2026, a table of ranks 1 to 11 with an agent column, where Kimi K3.1 Pro sits 10th at 59.7% in Claude Code.
The Terminal-Bench 4.1 leaderboard on 20 November 2026. Each row names the agent it ran in; the best open-weight model, Kimi K3.1 Pro, sits 10th at 59.7%.

Open weights win one board of four. tbench.ai's rows also mix agents (Claude Code, Codex, GitHub Copilot, Antigravity CLI and mini-swe-agent among them), so read its order as model plus agent.

Open weights give you control: the model runs on hardware you choose, your code stays there, and a licence file sets the terms. The cost is the hardware or a host, the setup, and on three of these four boards a lower score. If a subscription is the better buy for you, compare what Claude Code, Cursor and Copilot cost today. For hosted coding agents, my Google Jules alternatives post compares five.

A note from the person who builds Gravity. I build Gravity, the site you are reading, so a plain word on fit: Gravity is not a model or a coding tool, and it does not host or run any model on this list. It runs ready-made agents for jobs like research, lead generation and job search. The one that touches open source is the Open Source Contributor Guide, an agent that finds open source projects to contribute to: it reads your GitHub repositories, ranks active projects that use the same skills and welcome newcomers, and picks starter issues for you. It does not write your code. Gravity is in public beta, and as of November 2026 the plans are Free at $0 with no card, Mini Pilot at $15 a month, Auto Pilot at $20 and Auto Pilot Max at $100 (pricing).

FAQ

What is the best open source coding model right now?

Kimi K3.1 Pro, on the boards I read on 20 November 2026. It scores 67.2% on Artificial Analysis's Terminal-Bench 4.0, the top score on that chart, and ranks first of 63 open models on Arena's WebDev board at 1,683. If you need an MIT licence, pick DeepSeek V4.1 Flash 1017, the best open model on FrontierCode-AA 2.0 at 28.5%.

What is the best local LLM for coding?

Gemma 4, in its 26B A4B or 31B size, is the only pick here whose maker says it runs on consumer hardware: Google says quantized versions run natively on consumer GPUs, and Ollama's downloads are 16GB and 19GB. Qwen3.8-27B rates higher on Arena's open WebDev board (1,589 against 1,370) and is an 18GB Ollama download, but its card gives no hardware line. Small models are far weaker on long agentic tasks: Qwen3.8-27B scores 5.6% on Artificial Analysis's Terminal-Bench 4.0.

Can I run the top open models on a laptop?

Not the five big ones. Kimi K3.1 Pro, GLM-5.3, MiMo-V2.6-Pro, DeepSeek V4.1 Flash 1017 and Inkling 2 Small have weight files from 393.5 GB to 1,561 GB, and Inkling's card asks for a GPU cluster with at least 500GB of aggregated VRAM. Of the seven, only Gemma 4 comes with a maker's statement about consumer hardware.

Is an open model as good as Claude or GPT for coding?

It depends on the board. On Artificial Analysis's Terminal-Bench 4.0 the top open model leads, 67.2% against 65.2% for Gemini 4 Carbon (High), while on FrontierCode-AA 2.0 the best open model trails by about 20 points, 28.5% against 49.1%. On Arena's WebDev board Kimi K3.1 Pro ranks 17th overall, and on tbench.ai's Terminal-Bench 4.1 it ranks 10th, all read 20 November 2026.

Which open coding models allow commercial use?

All seven grant commercial use. DeepSeek V4.1 Flash 1017 and MiMo-V2.6-Pro (MIT) and Inkling 2 Small, Qwen3.8-27B and Gemma 4 (Apache 2.0) use OSI-approved licences, while Kimi K3.1 Pro and GLM-5.3 add conditions for model-hosting businesses above $20 million and $10 billion in revenue over 12 months. This is not legal advice: read the licence file of the exact repository you download.

What is the difference between open source and open weights?

The Open Source Initiative's Open Source AI Definition 1.0 asks for data information, code and parameters under OSI-approved terms. Downloadable weights alone are open weights. I checked licences only, so this post calls all seven open-weight models and says which use an OSI-approved licence.

Which benchmark should I trust for coding models?

Trust one that runs every model in the same harness and names its version. Artificial Analysis's Terminal-Bench 4.0 runs all 66 tasks with mini-swe-agent for every model, and FrontierCode-AA 2.0 grades whether a maintainer would merge the change. SWE-bench Verified has no entries after February 2026, so a November ranking built on it is out of date.

Is Qwen Coder any good?

Qwen's newest dedicated coder on Hugging Face, Qwen3-Coder-Next (Apache 2.0, 80B total and 3B active, from January 2026), is not on Arena's open WebDev list or Artificial Analysis's Terminal-Bench 4.0 chart as of 20 November 2026, so there is no current board score to judge it by. Qwen3-Coder-480B-A35B-Instruct sits 56th of 63 on Arena's open board at 1,274, while Qwen's general Qwen3.8-27B sits 13th at 1,589.

Is DeepSeek safe to use for code?

The weights are MIT-licensed, and if you run them yourself, the model runs on hardware you control. I did not review DeepSeek's hosted app or API terms, so read them before you send code to DeepSeek's own service.

Sources

  1. Artificial Analysis, Terminal-Bench 4.0 evaluation, read 20 November 2026 (the page states no update date; its chart shows 28 of 281 models).
  2. Artificial Analysis, FrontierCode-AA 2.0 evaluation, read 20 November 2026.
  3. Artificial Analysis, open-source models (parameters, context, open-weights flags), read 20 November 2026.
  4. Arena, Code Arena WebDev leaderboard with the Open Source filter, and the same board unfiltered, read 20 November 2026 (the page's stamp reads 19 or 20 November depending on the browser's time zone).
  5. Terminal-Bench 4.1 leaderboard, tbench.ai, read 20 November 2026 (its date column is each model's release date).
  6. SWE-bench leaderboards (newest Verified entry dated 26 February 2026), Aider LLM leaderboards (newest run 3 October 2025) and LiveCodeBench (results file last modified 1 August 2025), read 20 November 2026.
  7. Open Source Initiative, Open Source AI Definition 1.0 and OSI Approved Licenses, read 20 November 2026.
  8. Moonshot AI, Kimi K3.1 Pro model card, Kimi K3.1-Pro License and release post, read 20 November 2026.
  9. DeepSeek, DeepSeek V4.1 Flash 1017 model card and the DeepSeek V4.1 Flash base card, read 20 November 2026.
  10. Z.ai, GLM-5.3 model card and licence, GLM-5.3 docs, GLM-5.3-Flash card, and the licence tags on GLM-5.1 and GLM-5.2, read 20 November 2026.
  11. Xiaomi, MiMo-V2.6-Pro-RL card, MiMo-V2.6-Flash-RL card and release page, read 20 November 2026.
  12. Thinking Machines, Inkling 2 Small model card and release post (dated 5 November 2026, updated 12 November), read 20 November 2026.
  13. Alibaba Qwen, Qwen3.8-27B model card, Qwen3.8-Flash-Next card (Qwen Community License 1.0) and Qwen3-Coder-Next card, read 20 November 2026.
  14. Google, Gemma 4 31B and Gemma 4 26B A4B model cards and the Gemma 4 launch post (dated 2 April 2026), read 20 November 2026.
  15. Mistral, Mistral Large 4 and Mistral Large 3 model cards (licence tags), read 20 November 2026.
  16. Ollama library pages for qwen3.8 and gemma4 (the runner's download sizes), read 20 November 2026.
  17. Screenshots captured 20 November 2026 at 1440 pixels wide, logged out. The Artificial Analysis shot was cropped by 16 pixels at the top to remove a cut-off line of text; nothing else was edited.
Share