Ornith-1.5 Review: The Self-Improving Open Model That Matches Claude Opus 4.8
What is Ornith-1.5?
On August 19, 2026, DeepReinforce released Ornith-1.5 — an open-weight model family under the MIT license that arrives in three sizes:
- 397B MoE flagship
- 35B MoE (3B active per token)
- 9B dense with a mobile-quantized build for iPhone and Android
The headline isn't the size — it's the training method. Ornith extends "self-scaffolding" into a full self-improvement loop: the model proposes its own tasks, builds scaffolds for them, and generates solution rollouts that become the reinforcement-learning signal. No fixed human-authored task set.
Benchmarks (vendor-reported)
| Model | Terminal-Bench 2.1 | SWE-bench Verified | Notes |
|---|---|---|---|
| Ornith-1.5-397B | 86.1 | 86.0 | Matches Claude Opus 4.8 (85.0) |
| GLM-5.3 | 28.3 (TB 3.0) | — | Best open coding (GLM-5.3 review) |
| Qwen3.8-27B | 73.0 (TB 2.1) | 61.7 (Pro) | Local-runnable (Run Qwen3.8-27B locally) |
| Kimi K3 | 88.3 | — | 2.8T, leads open field |
- Context: 262,144 tokens (YaRN to ~1M)
- License: MIT, in BF16 / FP8 / NVFP4 / GGUF / MLX
- Availability: weights on Hugging Face day one; vLLM, Ollama, OpenCode support
Caveat: these are Ornith's own runs, averaged over five attempts, harness published per benchmark. Independent reproduction is still landing — treat the Opus-4.8 parity as claimed, not yet confirmed.
How to run it locally
- 9B: runs on a single GPU; the
-Mobilequant targets phones (sub-8GB). - 35B-A3B: the deployment sweet spot — outperforms Gemma-4-31B and Muse-Glimmer-30B on agentic coding while activating only 3B/token.
- GGUF / MLX builds are day-one on Hugging Face; serve via
llama.cpp,ollama, orvllm.
How it compares for builders
Ornith-1.5 matters if you want open-weight agentic coding at near-frontier quality without a closed-model bill. The 35B is the one to actually deploy. For lighter local work, Qwen3.8-27B (Run Qwen3.8-27B locally) is the more practical pick — smaller, faster, same license family.
Bottom line
Ornith-1.5 is the most interesting open release of late August 2026 — not because it's biggest, but because the self-improvement loop is a genuine methodological bet. Wait for independent benchmarks, but the 35B is worth testing now.
Related: AI Agent Guide · GLM-5.3 review · Run Qwen3.8-27B locally · OrcaRouter
Next Steps
Compare the other open coding models in this cluster, then pick a local deploy path that actually fits your hardware.
Read the GLM-5.3 review →Run Qwen3.8-27B locally →Create your own AI agent for free →how do AI agents work — return to the complete AI agent architecture guide.
Was this helpful?
Your feedback stays on this page — no tracking.