DeepReinforce has released Ornith-1.5, a new family of open AI models designed around a self-improving training approach in which the models generate not only solutions but also the tasks and scaffolds used to train themselves. The new release comes in three variants: a 397-billion-parameter mixture-of-experts (MoE) flagship, a 35B MoE model that activates only 3B parameters per token, and a 9B dense model that also ships with a quantized mobile build for iPhone and Android.
The release builds on Ornith-1.0’s idea of treating the AI scaffold — the instructions, tools and orchestration used to solve a task — as something the model can learn. Ornith-1.5 takes that concept further by creating a closed self-improvement loop: the system proposes increasingly difficult tasks, generates task-specific scaffolds, produces solution rollouts and then uses reinforcement learning to improve all three components. DeepReinforce says the new models show strong results across coding, reasoning and agentic benchmarks, with the flagship reaching 85.1 on Terminal-Bench 2.1 and 56.0 on DeepSWE.
Ornith-1.5 Comes In Three Model Sizes
The new release is designed to cover different levels of computing requirements.
At the top is Ornith-1.5-397B, a mixture-of-experts model aimed at frontier-scale workloads. The 35B version is also an MoE model but activates only 3 billion parameters per token, potentially reducing inference requirements compared with running all 35 billion parameters. The smallest 9B model is a dense model and includes a quantized mobile version intended for smartphones.
Ornith-1.5 Model Lineup
| Model | Architecture | Total Parameters | Active Parameters | Key Positioning |
|---|---|---|---|---|
| Ornith-1.5-397B | MoE | 397B | Not specified in source | Flagship / frontier workloads |
| Ornith-1.5-35B | MoE | 35B | 3B/token | Efficient high-performance model |
| Ornith-1.5-9B | Dense | 9B | 9B | Smaller deployment |
| 9B Mobile Build | Quantized | 9B | — | iPhone + Android |
The range gives developers an option to choose between maximum capability and easier deployment.
From Self-Scaffolding To Self-Improvement
The biggest technical change from Ornith-1.0 is the move from self-scaffolding to a more complete self-improvement loop.
In the previous generation, the model could generate scaffolds around a fixed collection of human-curated tasks. Ornith-1.5 allows the system to propose the tasks themselves and then build a suitable scaffold for each task.
How Ornith-1.5 Trains Itself
Existing Environment / Codebase
↓
Model Proposes A New Task
↓
Task-Specific Scaffold Generated
↓
Model Produces Solution Rollout
↓
Reward Evaluates Task + Scaffold + Solution
↓
Reinforcement Learning
↓
Better Tasks + Better Scaffolds + Better Solutions
This creates a feedback loop in which the training system is not limited to improving the final answer.
The Model Looks For Tasks Near Its Capability Limit
Ornith-1.5 does not simply generate random tasks.
According to DeepReinforce’s published methodology, the system attempts to generate progressively harder problems that sit near the model’s current capability frontier. The target is an empirical success rate of approximately 20%. Once the model begins solving a task reliably, the task becomes less valuable to the generator.
Difficulty Selection
| Task Outcome | Training Value |
|---|---|
| Too easy | Lower value |
| Around 20% success | Target difficulty |
| Frequently solved | Lower value |
| Invalid / unverifiable | Zero reward |
| Novel and challenging | Higher potential value |
The approach is designed to prevent the training process from spending too much effort on problems the model has already mastered.
Three Signals Determine Task Reward
The reward mechanism considers three major signals.
First, the task and scaffold need to create a valid and verifiable environment. Second, the difficulty should sit close to the model’s current capability frontier. Third, the task should be sufficiently novel compared with problems already generated.
Validity acts as a hard gate, meaning malformed or unverifiable tasks receive no reward. All three stages are optimised using GRPO, a reinforcement-learning approach used in the training process.
Ornith-1.5 Reward System
| Reward Signal | Purpose |
|---|---|
| Validity | Ensures task can actually be evaluated |
| Frontier difficulty | Keeps tasks challenging |
| Novelty | Avoids repeatedly generating similar problems |
| GRPO | Optimises the three-stage process |
This structure attempts to address a major challenge in self-improving AI: preventing the model from creating training data that looks useful but cannot reliably measure progress.
Ornith-1.5 Shows Strong Coding Benchmark Results
DeepReinforce’s published results put the 397B model near leading proprietary models on some coding benchmarks.
Across five independent runs, the company reports 85.1 on Terminal-Bench 2.1 and 56.0 on DeepSWE for Ornith-1.5-397B. DeepReinforce compares these results with Claude Opus 4.8 at 85.0 and 59.0 respectively.
Flagship Benchmark Comparison
| Model | Terminal-Bench 2.1 | DeepSWE |
|---|---|---|
| Ornith-1.5-397B | 85.1 | 56.0 |
| Claude Opus 4.8 | 85.0 | 59.0 |
| GLM-5.2 | Lower, according to source | Lower, according to source |
| DeepSeek-V4-Flash-0731 | Lower, according to source | Lower, according to source |
The comparison is based on DeepReinforce’s published tables and should therefore be treated as company-reported benchmark results, rather than an independent evaluation.
Ornith-1.5 397B Benchmark Snapshot
Terminal-Bench 2.1Ornith-1.5-397B █████████████████ 85.1Claude Opus 4.8 █████████████████ 85.0DeepSWEClaude Opus 4.8 ████████████ 59.0Ornith-1.5-397B ███████████ 56.0
The results show a particularly close score on Terminal-Bench, while Claude Opus 4.8 remains ahead on DeepSWE.
The 35B Model Activates Only 3B Parameters Per Token
One of the more interesting models in the lineup is Ornith-1.5-35B.
Although the model has 35 billion total parameters, it activates only 3 billion parameters per token during inference. This MoE design can potentially reduce the computational cost of running the model while retaining a much larger total parameter pool.
The model reportedly scores 68.5 on Terminal-Bench 2.1 and 79.0 on SWE-Bench Verified.
Ornith-1.5-35B
| Specification | Figure |
|---|---|
| Total parameters | 35B |
| Active parameters/token | 3B |
| Activation ratio | ~8.6% |
| Terminal-Bench 2.1 | 68.5 |
| SWE-Bench Verified | 79.0 |
| Architecture | MoE |
Only around 8.6% of the model’s total parameters are activated per token, based on the reported 3B active parameters out of 35B total.
The 9B Model Brings Ornith To Smaller Devices
Ornith-1.5 also includes a 9B dense model, giving developers a substantially smaller option than the 397B flagship.
The 9B version reaches 47.0 on Terminal-Bench 2.1 and 70.6 on SWE-Bench Verified, according to DeepReinforce’s published results. The company places the model above Gemma 4-31B and Qwen 3.6-35B in its reported comparisons.
More notably, a quantized mobile build is shipping for iPhone and Android, potentially making the model usable in much more resource-constrained environments.
9B Model Performance
| Benchmark | Ornith-1.5-9B |
|---|---|
| Terminal-Bench 2.1 | 47.0 |
| SWE-Bench Verified | 70.6 |
| Deployment | Mobile quantized build |
| Platforms | iPhone + Android |
This gives Ornith a model that can target local and mobile use cases rather than being restricted to large data-centre deployments.
Ornith-1.5 Goes Beyond Coding
Although coding is a major focus of the release, the model family is also designed for reasoning and agentic tasks.
At flagship scale, Ornith-1.5-397B reportedly scores 92.8 on GPQA Diamond and 86.6 on BrowseComp.
Broader Benchmark Coverage
| Capability | Benchmark | Reported Score |
|---|---|---|
| Coding | Terminal-Bench 2.1 | 85.1 |
| Coding | DeepSWE | 56.0 |
| Coding | SWE-Bench Verified | 79.0 for 35B |
| Reasoning | GPQA Diamond | 92.8 |
| Browsing / agentic work | BrowseComp | 86.6 |
This suggests DeepReinforce is positioning Ornith-1.5 as a broader agentic model rather than a coding-only system.
Ornith-1.5 Builds On Ornith-1.0
The new release follows Ornith-1.0, which DeepReinforce open-sourced in June 2026.
Ornith-1.0 was released in 9B dense, 31B dense, 35B MoE and 397B MoE variants under an MIT license. The models were post-trained on Gemma 4 and Qwen 3.5 checkpoints and introduced the concept of making the scaffold itself a learnable component.
Ornith Evolution
| Generation | Key Development |
|---|---|
| Ornith-1.0 | Self-scaffolding |
| Ornith-1.5 | Self-generating tasks + scaffolds + rollouts |
| Future direction | More autonomous improvement |
The move from 1.0 to 1.5 therefore represents a change in how the models are trained, not simply an increase in model size.
Ornith-1.0 Already Targeted Agentic Coding
The original Ornith-1.0 release was designed around agentic coding and self-generated reinforcement-learning scaffolds.
DeepReinforce reported that its 397B model reached 77.5 on Terminal-Bench 2.1 and 82.4 on SWE-Bench Verified, which the company said was competitive with Claude Opus 4.7 and other leading open models at the time.
Ornith 1.0 Vs 1.5
| Metric | Ornith-1.0 397B | Ornith-1.5 397B |
|---|---|---|
| Terminal-Bench 2.1 | 77.5 | 85.1 |
| Improvement | — | +7.6 points |
| Training approach | Self-scaffolding | Closed self-improvement loop |
The reported Terminal-Bench score increased by 7.6 points between the two generations.
Terminal-Bench 2.1Ornith-1.0 ███████████████ 77.5Ornith-1.5 █████████████████ 85.1
The improvement is based on DeepReinforce’s published benchmark results for the two releases. It should not be interpreted as a controlled apples-to-apples scientific comparison unless the evaluation conditions are identical.
Open Models Could Put Pressure On Proprietary AI
One of Ornith-1.5’s biggest implications is its availability as an open model family.
Open models allow developers and companies to download weights, customise models and potentially deploy them on their own infrastructure.
This can be particularly important for businesses that do not want to send sensitive code or data to external AI providers.
Proprietary Vs Open Deployment
| Factor | Proprietary AI | Open Model |
|---|---|---|
| Model weights | Usually unavailable | Available |
| Customisation | Provider-dependent | Greater flexibility |
| Deployment | Provider infrastructure | Can run privately |
| Data control | Depends on provider | Potentially greater |
| Cost structure | API / subscription | Infrastructure-dependent |
| Developer control | Limited | Higher |
The trade-off is that running very large open models can require substantial computing resources.
The 397B Model Is Not Designed For Ordinary Hardware
Although Ornith is open, that does not mean every model can run on a consumer laptop.
The 397B MoE flagship is designed for large-scale workloads and requires significant infrastructure.
The 35B and 9B versions provide more practical alternatives, while the quantized 9B mobile build is the most accessible option in the lineup.
Model Size Vs Accessibility
397B ████████████████████ Frontier / Data Centre35B ████████ Advanced Local / Server9B ██ Smaller Deployment9B Q ██ Mobile
The model family therefore covers a broad range from frontier-scale AI infrastructure to mobile deployment.
Self-Generated Training Data Could Change AI Development
The most important part of Ornith-1.5 may ultimately be its training methodology.
AI companies have traditionally relied heavily on human-created datasets, synthetic data pipelines and manually designed evaluation environments.
Ornith-1.5 attempts to make the model itself part of the process that creates new training challenges.
Traditional AI Training
Human Tasks
↓
Human Scaffolds
↓
AI Solutions
↓
Reward
Ornith-1.5
AI Generates Task
↓
AI Generates Scaffold
↓
AI Generates Solution
↓
Reward
↓
AI Improves All Three
If this approach scales effectively, models could potentially generate increasingly challenging training environments without requiring humans to manually design every new task.
Reward Hacking Remains A Challenge
DeepReinforce has previously researched reinforcement-learning optimisation and has acknowledged reward hacking as a concern in its work.
Ornith-1.5 attempts to address part of this problem by requiring generated tasks to be valid and verifiable. An invalid task receives no reward.
That is important because a self-improving system could otherwise learn to exploit weaknesses in its own evaluation process.
Self-Improvement Safety Mechanism
Generate Task
↓
Check Validity
↓
Check Difficulty
↓
Check Novelty
↓
Generate Rollout
↓
Reward
↓
Update Model
The approach does not eliminate all risks associated with automated reinforcement learning, but it creates explicit checks around the quality of generated training environments.
The Open-Source Release Is Available On Hugging Face
DeepReinforce has released the Ornith-1.5 model family through its Hugging Face collection. The release includes the three major model sizes described in the announcement.
Developers interested in experimenting with the models can access the weights and related materials through the project’s Hugging Face collection.
What Ornith-1.5 Means For AI Coding
The release comes as AI coding agents are moving from simple autocomplete toward systems capable of planning, using tools, executing code and completing multi-step software tasks.
Ornith-1.5’s training methodology is specifically designed around this type of agentic workflow.
The combination of self-generated tasks and scaffolds could potentially help the models improve on tasks that require more than simply predicting the next line of code.
Agentic Coding Loop
Understand Task
↓
Plan
↓
Use Tools
↓
Write Code
↓
Run Tests
↓
Fix Errors
↓
Complete Task
The benchmark results indicate that coding remains one of Ornith-1.5’s primary strengths, although real-world performance will depend on deployment environment, tool access and the complexity of individual tasks.
The Bigger Picture
Ornith-1.5 is notable not simply because DeepReinforce has released another large open model, but because it pushes the idea of self-improving AI training further. The system generates increasingly difficult tasks, creates the scaffolds needed to solve them and produces solution rollouts, with reinforcement learning used to improve all three stages.
The three-model lineup also makes the release unusually broad. The 397B flagship targets frontier-scale workloads, the 35B MoE activates only 3B parameters per token, and the 9B model is small enough to have a quantized mobile version for iPhone and Android. This gives developers options ranging from large-scale infrastructure to local and mobile deployments.
Looking Ahead
The biggest test for Ornith-1.5 will be whether its self-improvement approach translates into sustained gains beyond the benchmarks published by DeepReinforce. Generating increasingly difficult and genuinely useful tasks is considerably harder than simply producing more synthetic data. The validity, novelty and difficulty controls built into the training loop will therefore be important as the system scales.
For developers, the open release gives Ornith-1.5 a chance to compete with proprietary coding and reasoning systems while offering greater control over deployment. The 397B model targets high-end infrastructure, while the 35B and 9B versions could make the technology accessible to a wider range of developers. If the reported benchmark gains hold up under independent testing, Ornith-1.5 could become another important example of how open models are narrowing the gap with leading proprietary AI systems
Get the day’s top stories in your inbox
One concise email. No spam, unsubscribe anytime.
