DeepReinforce launches Ornith-1.0, an open-source coding MoE trained with GRPO that scores 82.4 on SWE-bench Verified

Original post

Aloha! 🌺 Meet Ornith-1.0, a family of open-source LLMs specialized for agentic coding.

Ornith-1.0 spans the full parameter sizes including 9B Dense, 31B Dense, 35B MoE, and 397B MoE. It achieves state-of-the-art performance among open-source models of comparable size on coding benchmarks including: ✅Terminal-Bench 2.1(77.5) ✅SWE-Bench(82.4 on verified, 62.2 on pro, 78.9 on Multilingual) ✅NL2Repo(48.2) ✅SWE Atlas(41.2 on QnA, 42.6 RF, 39.1 TW) ✅ClawEval(77.1)

Post-trained on top of gemma4 and qwen3.5, Ornith-1.0 employs a novel self-improving training strategy in which reinforcement learning is used to generate not only solution rollouts, but also the task-specific scaffolds that drive those rollouts. By jointly optimizing the scaffold and the resulting solution, the model generate higher-quality solutions in agentic coding.😎

All models are released under the MIT license, enabling full commercial and research use.

📖Tech Blog: http://deep-reinforce.com/ornith_1_0.html 🤗Huggingface: http://huggingface.co/collections/deepreinforce-ai/ornith-10

7:15 AM · Jun 25, 2026 · 769K Views

VIEWS4.2KBOOKMARKS11LIKES54RETWEETS5REPLIES5

Ornith@ornith_

Ornith-1.0-35B strikes a strong balance between capability and efficiency, achieving 64.2 on Terminal-Bench 2.1 and 75.6 on SWE-Bench Verified. It delivers best-in-class performance among 35B-class models, outperforming Qwen 3.5-35B, Qwen 3.6-35B, and Gemma 4-31B, while even surpassing Qwen 3.5-397B on Terminal-Bench 2.1.

Huggingface: http://huggingface.co/deepreinforce-ai/Ornith-1.0-35B

4h4.2K5411

Ornith@ornith_

Ornith-1.0-9B is easily deployable on edge devices, achieves best-in-class performance among models of similar size (43.1 for Terminal-Bench 2.1 and 69.4 for SWE Verified ), and matches or exceeds the performance of much larger models, including Gemma 4-31B and Qwen 3.6-35B.

Huggingface: https://huggingface.co/deepreinforce-ai/Ornith-1.0-9B

4h3.7K537

Ornith@ornith_

Ornith-1.0 9B and 35B deliver great performance. We've also made GGUF versions, so you can easily run them in http://atomic.chat @atomic_chat_hq, Ollama @ollama and Unsloth @UnslothAI. All models are open-source at HuggingFace @huggingface

Link: http://huggingface.co/collections/deepreinforce-ai/ornith-10 Github: http://github.com/deepreinforce-ai/Ornith-1 Blog: http://deep-reinforce.com/ornith_1_0.html

4h3K408

AshutoshShrivastava@ai_for_success

@ornith_ open-source must win, Can't wait to run this locally and check this tomorrow .

4h1.3K14

Rohan Paul@rohanpaul_ai

@ornith_ So great to see reward hacking treated as a training problem, not an afterthought.

3h9336

Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@teortaxesTex

almost all engagement with this seems synthetic I'd be interested in serious people like @_xjdr giving it a spin, until then I reserve my skepticism that Qwen 3.5 base can approach or surpass GLM 5.2 just like that

Ornith@ornith_

Aloha! 🌺 Meet Ornith-1.0, a family of open-source LLMs specialized for agentic coding.

All models are released under the MIT license, enabling full commercial and research use.

📖Tech Blog: http://deep-reinforce.com/ornith_1_0.html 🤗Huggingface: http://huggingface.co/collections/deepreinforce-ai/ornith-10

8m39321

s.h@hebbarmp

@ornith_ DeepSWE?

3h6049

Przemysław Skrzypek@brzewVCE

@ornith_ Looking forward to this, but the 31b dense is missing

2h15911

Pawzard@pawzzard

@ornith_ named it after yourself and dropped four model sizes

bro pulled up to the open source cookout like its your going away party

2h6647

Charlotte Xia@xia_char

@ornith_ @deep_reinforce Goooo Ornith!!

3h6792

benjamin@newondiscord

@ornith_ smells like benchmaxxing

2h4355

Alberto Fuentes (e/acc)@AlberFuen

@ornith_ How many active experts the Moe?

3h359

Rohan Paul@rohanpaul_ai

Ornith is trained to improve the task scaffold and the code solution together, so it learns the workflow that produces better answers, not just the answer itself.

The reward signal goes back into both parts, so over many RL steps it can learn better planning, retry logic, tool use, and error handling for coding-agent tasks.

📖Tech Blog: http://deep-reinforce.com/ornith_1_0.html

🤗Huggingface: http://huggingface.co/collections/deepreinforce-ai/ornith-10

Rohan Paul@rohanpaul_ai

Another fantastic open source release.

DeepReinforce just dropped Ornith-1.0, an MIT-licensed open-source family of agentic coding LLMs.

The flagship Ornith-1.0-397B MoE (17B-active) is the most powerful model in the release, reporting 82.4 on SWE-Bench Verified and 77.5 on Terminal-Bench 2.1 - surpassing Claude Opus 4.7 on both benchmarks.

Built on top of pretrained Gemma 4 and Qwen 3.5

Employs a novel self-improving training strategy. With this Ornith changes the training target by asking the model to improve both the answer and the task scaffold, meaning the plan, memory pattern, tool rhythm, error handling, and search process that shape the answer.

During RL, the model proposes a better scaffold first, then uses it to produce solution rollouts, and the reward updates both stages together.

That makes the model less like a coder following one rigid checklist and more like a coder learning which checklist works for each type of bug, repo, or terminal task.

The most interesting result is the 9B model reaching 69.4 on SWE-Bench Verified

1h59630

Tabib hasan@Tabib_Hasan_

@ornith_ @grok can you compare Ornith-1.0’s reported scores against the current open-source leaders on Terminal-Bench and SWE-Bench Verified, and point out which numbers are author-reported vs independently listed on public leaderboards? and also compare with top frontier models

1h174