header-langage
简体中文
繁體中文
English
Tiếng Việt
한국어
日本語
ภาษาไทย
Türkçe
Scan to Download the APP
Partnered With:

Li Jiwei Open Sources Ornith Programming Model: 397B Flagship Edition Beats Claude Opus 4.7 in Multiple Programming Benchmarks

According to Perceive Beating monitoring, the AI team DeepReinforce founded by renowned NLP scholar Ji Li has open-sourced four models specializing in programming agents under the Ornith-1.0 release, all licensed under the MIT license. Ornith-1.0 is trained on Gemma 4 and Qwen 3.5 and encompasses two architectures: Dense and MoE (Mixture of Experts). In the Terminal-Bench 2.1 (Claude Code) programming agent test, the flagship Ornith-1.0-397B MoE achieved a score of 78.2, surpassing the closed-source model Claude Opus 4.7 at 69.7.


The core of Ornith-1.0 lies in the joint reinforcement learning of the agent's runtime environment scaffold and the policy model responsible for code generation. Unlike the traditional approach of a fixed development environment, the model spontaneously designs and updates the scaffold for the current task during training (e.g., custom memory logging and error-handling retry logic), then generates code to run in the new environment. The reinforcement learning reward signal is simultaneously propagated back to optimize both stages.


Due to the model's susceptibility to reward hacking during self-evolving scaffold creation, such as by reading hidden test files or manipulating test scripts, the development team deployed a three-tier security defense: in addition to physically sandboxing external tools and testing arena boundaries, the system uses a deterministic monitor to intercept unauthorized access behavior. Furthermore, a frozen third-party large model serves as a monitoring referee, exercising veto power over solutions that comply with the rules but exhibit intent to game the system.


Evaluation data indicates the model's scalability. In the same Claude Code testing environment, the 35B MoE model Ornith-1.0-35B achieved a score of 62.8, surpassing the 397B large-parameter model Qwen 3.5-397B (48.6) and Qwen 3.6-35B (49.2). The edge-specific Ornith-1.0-9B Dense model also scored 69.4 in the SWE-Bench Verified test, outperforming the 31B Gemma 4-31B (52.0).

举报 Correction/Report
Correction/Report
Submit
Add Library
Visible to myself only
Public
Save
Choose Library
Add Library
Cancel
Finish