LLM Releases

Coding and software agents

All reports

Open coding model releases

A focused comparison of open-weight or open-source models positioned for coding, repository work, tool use, and software-engineering agents.

Open-weight coding models have closed much of the gap with proprietary frontier systems. This table compares the downloadable ones built for software work, ranked by their SWE-bench Verified claim — resolving real GitHub issues — then by recency.

Size and context matter as much as the headline score: a smaller model with a long context can be the better fit for whole-repository work on local hardware.

SWE-bench values are included only when the catalog has a sourced claim or verified entry. Vendor-reported scores remain claims until independently verified.

ModelLabLicenseParamsContextSWE-benchReleasedSource
Nex-N2-ProNex AGIApache-2.0397B · 17B active262K80.8%Jun 2, 2026source
Kimi K2.6MoonshotModified MIT1T · 32B active256K80.2%Mar 30, 2026source
Hunyuan Hy3HunyuanApache-2.0295B · 21B active256K78%Jul 6, 2026source
GLM-5Z.aiMIT744B · 40B active—77.8%Feb 11, 2026source
Mistral Medium 3.5MistralMistral Research / Commercial128B256K77.6%Mar 18, 2026source
Qwen3.6-27BQwenApache-2.027B256K77.2%May 12, 2026source
Kimi K3MoonshotModified MIT2.8T · 50B active1.0M76.8%Jul 16, 2026source
Kimi K2.5MoonshotModified MIT1T · 32B active256K76.8%Jan 27, 2026source
GLM-4.7Z.aiMIT358B · 32B active—73.8%Jan 8, 2026source
Qwen3-Coder-NextQwenApache-2.080B · 3B active262K70.6%Feb 3, 2026source
Sarvam-105BSarvamApache-2.0105B · 10.3B active128K45%Mar 6, 2026source
Agnes 3.0 FlashAgnes AIApache 2.0 (preview checkpoint)33B262K-Sep 14, 2026source
Atria Dawn PreviewShanghai AI LabMIT744B262K-Sep 11, 2026source
Ling-3.0-flash-VLinclusionAIMIT124B · 5.5B active262K-Sep 10, 2026source
DeepSeek-V4.1-FlashDeepSeekMIT552B · 16B active1M-Sep 10, 2026source
Nex-N2.5-MiniNex AGIApache-2.035B · 3B active262K-Sep 8, 2026source
Nex-N2.5-MaxNex AGIApache-2.01.6T · 49B active1.0M-Sep 8, 2026source
Nex-N2.5-ProNex AGIApache-2.0397B · 17B active262K-Sep 8, 2026source
MiniCPM5-2BOpenBMBApache 2.02.52B131K-Sep 7, 2026source
LLaDA2.2-miniinclusionAIApache-2.016B · 1.4B active131K-Sep 5, 2026source
Ling-3.0-flash-SanteinclusionAISante-specific weights not confirmed posted at launch (hosted API-first); base Ling-3.0-flash family is MIT124B · 5.1B active262K-Sep 4, 2026source
K2 Horizon 375B A23BMBZUAIApache-2.0375B · 23B active524K-Sep 3, 2026source
Muse Spark 1.3MetaMeta license (weights pending)Undisclosed1.0M-Sep 2, 2026source
Hy4 previewHunyuanApache-2.0770B · 49B active1.0M-Aug 28, 2026source
Ling-3.0-flash-FininclusionAIMIT124B · 5.1B active262K-Aug 27, 2026source
Qwen3.8-Flash-NextQwenQwen Community License 1.0125B · 6B active262K-Aug 26, 2026source
GLM-5.3-FlashZ.aiMIT320B · 18B active1M-Aug 26, 2026source
Granite 4.2 3BIBMApache-2.04B131K-Aug 25, 2026source
Granite 4.2 8BIBMApache-2.09B131K-Aug 25, 2026source
Granite 4.2 30BIBMApache-2.029B131K-Aug 25, 2026source
Apodex 1.1 miniApodexApache 2.0 (open weights announced; rolling out at launch)35B—-Aug 24, 2026source
Ornith-1.5-9BDeepReinforceMIT9B—-Aug 19, 2026source
Ornith-1.5-35B-A3BDeepReinforceMIT35B · 3B active—-Aug 19, 2026source
Ornith-1.5-397BDeepReinforceMITUndisclosed—-Aug 19, 2026source
GLM-5.3Z.aiGLM-5.3 License (MIT-style; security-review gate for >$10B-revenue Model-as-a-Service providers)743B1.0M-Aug 14, 2026source
Qwen3.8-27BQwenApache-2.027B262K-Aug 14, 2026source
Dots3-Note PreviewDots StudioApache-2.0280B · 16B active512K-Aug 14, 2026source
North Micro Vision InstructCohereApache-2.02.4B—-Aug 12, 2026source
LFM2.5-VL-3BLiquidLFM Open License v1.03.1B—-Aug 12, 2026source
Nemotron 3.5 LightningNVIDIAOpenMDW-1.131.6B · 3.6B active1M-Aug 11, 2026source
Muse GlimmerMetaApache-2.030B131K-Aug 10, 2026source
LFM2.5-2.6BLiquidLFM Open License v1.02.69B131K-Aug 6, 2026source
Ling-3.0-tinyinclusionAIMIT7.9B · 1.3B active262K-Aug 6, 2026source
Muse Spark 1.2MetaMeta license (weights pending)Undisclosed1.0M-Aug 5, 2026source
Qwen3.8-MaxQwenApache-2.02.4T · 95B active1.0M-Aug 3, 2026source
DeepSeek-V4-Flash-0731DeepSeekMIT284B · 13B active1M-Jul 31, 2026source
Ling-3.0-flashinclusionAIApache 2.0 (announced; weights not yet posted as of 2026-07-24)124B · 5.1B active262K-Jul 23, 2026source
Laguna S 2.1PoolsideOpenMDW-1.1118B · 8B active1M-Jul 21, 2026source
InklingThinking MachinesApache-2.0975B · 41B active1.0M-Jul 15, 2026source
Nemotron-Labs-3-Puzzle-75B-A9BNVIDIAOpenMDW-1.175.3B · 9.3B active1M-Jul 6, 2026source
Mistral frontier open-weight MoE (unnamed)MistralOpen weightsUndisclosed—-Jul 6, 2026source
Laguna XS 2.1PoolsideOpenMDW-1.133B · 3B active256K-Jul 2, 2026source
LongCat-2.0MeituanMIT1.6T · 48B active1M-Jun 30, 2026source
Kimi K2.7 CodeMoonshotModified MIT1T · 32B active262K-Jun 18, 2026source
GLM-5.2Z.aiMIT753B1M-Jun 17, 2026source
MiniMax-M3MiniMaxMiniMax Community License428B · 23B active1M-Jun 16, 2026source
DiffusionGemma 26B-A4BDeepMindApache-2.025.2B · 3.8B active256K-Jun 10, 2026source
North Mini Code 1.0CohereApache-2.030B · 3B active256K-Jun 9, 2026source
Nemotron 3 Ultra 550B-A55BNVIDIANemotron Open Model License550B · 55B active1M-Jun 4, 2026source
Gemma 4 12BDeepMindApache-2.012B256K-Jun 3, 2026source
Step-3.7-FlashStepFunApache-2.0196B · 11B active256K-May 29, 2026source
LFM2.5-8B-A1BLiquidLFM Open License v1.08.3B · 1.5B active131K-May 28, 2026source
MiniMax-M2.7MiniMaxMiniMax Model License229.9B · 9.8B active—-May 26, 2026source
DeepSeek V4-FlashDeepSeekMIT284B · 13B active1M-Apr 24, 2026source
DeepSeek V4-ProDeepSeekMIT1.6T · 49B active1M-Apr 24, 2026source
Hunyuan Hy3-previewHunyuanTencent License295B · 21B active256K-Apr 23, 2026source
Hunyuan-A13B-InstructHunyuanTencent Hunyuan A13B License80B · 13B active—-Apr 22, 2026source
MiMo-V2.5-ProXiaomiMIT1.0T · 42B active1M-Apr 22, 2026source
MiMo-V2.5XiaomiMIT310B · 15B active1M-Apr 22, 2026source
GLM-5.1Z.aiMIT754B · 40B active—-Apr 8, 2026source
Gemma 4 31BDeepMindApache-2.031B—-Apr 2, 2026source
Nemotron 3 Super 120B-A12BNVIDIANemotron Open Model License120B · 12B active1M-Mar 16, 2026source
Mistral Small 4MistralApache 2.0119B · 6B active256K-Mar 16, 2026source
Step-3.5-FlashStepFunApache-2.0196B · 11B active256K-Mar 14, 2026source
Qwen3.5-9BQwenApache-2.09B262K-Mar 2, 2026source
Qwen3.5-4BQwenApache-2.04B262K-Mar 2, 2026source
Qwen3.5-2BQwenApache-2.02B262K-Mar 2, 2026source
Qwen3.5-0.8BQwenApache-2.00.8B262K-Mar 2, 2026source
Qwen3.5-397BQwenApache-2.0397B · 17B active1M-Feb 20, 2026source
OLMo 3 Think 32BAi2Apache-2.032B—-Dec 15, 2025source

Frequently asked questions

What makes a model an “open coding model”?

It must be downloadable (open-weight or open-source) and positioned for software work — coding, repository tasks, tool use, or software-engineering agents. Models behind a closed API are covered in the frontier leaderboard instead.

Is SWE-bench the only thing that matters for coding?

No. SWE-bench Verified is the most comparable single number, but context window (how much of a repo fits), license terms for commercial use, and inference cost or local hardware fit often decide which model is actually usable for you.

Can I use these models commercially?

Often, but check each license. Open-source licences like Apache-2.0 and MIT are permissive; many open-weight models carry custom terms with scale, competitor, or redistribution restrictions. See the open-weight guide for a checklist.

Related