What is 2026 Ornith-1.5: self-improving 397B and 9B mobile weights

Ornith-1.5 shipped open weights on 19 August. This piece separates official numbers from GIGAZINE’s MIT / Hugging Face write-up.

On 19 August 2026 Ornith published three downloadable open-weight sizes: a large mixture-of-experts aimed at agentic coding scores that sit next to a closed flagship, and a dense 9B the authors say can be quantized onto phones.

The object is Ornith-1.5. The official blog frames it as end-to-end self-improvement: the model proposes harder tasks, writes a task-specific scaffold, then produces solution rollouts for reinforcement learning. What follows is only official figures plus GIGAZINE’s distribution notes — not a claim that an outside lab has reproduced a win over Opus.

What shipped

The post says 1.5 extends 1.0, and that 1.0 was built on Qwen3.5 and Gemma 4 with continued pretraining, mid-training and post-training. 1.5 moves the loop from optimizing scaffolds and rollouts to jointly optimizing task generation as well.

Figures that can be checked

1. 1

   The three sizes are named

   The official list is Ornith-1.5-397B (MoE), Ornith-1.5-35B-A3B (MoE, 3B active per token), and Ornith-1.5-9B (dense). GIGAZINE also named the quant pack Ornith-1.5-9B-Mobile for iPhone and Android.

2. 2

   397B versus Opus 4.8 is mixed, not a sweep

   The post reports 86.1 on Terminal-Bench 2.1 and 56.0 on DeepSWE, “on par” with Claude Opus 4.8 at 85.0 and 59.0. On the same table Terminus-2 is above Opus and DeepSWE is below. Against GLM-5.2 and DeepSeek-V4-Flash-0731 the authors claim a lead among open models of similar scale.

3. 3

   The curriculum is generated; reward has a hard gate

   Each cycle proposes harder tasks from an environment, task-type instructions and prior history; writes a scaffold (instructions, tools, decomposition, orchestration); then produces rollouts. Task reward is validity V × frontier difficulty D × novelty N. If V=0 the whole term is zero. Difficulty uses the current model’s success rate with target p*=0.2. All three stages use GRPO.

The official loop is described as: stronger policies yield harder, more informative tasks; evolving scaffolds elicit more capability; higher-quality rollouts supply the next learning signal. That is a methods claim, not an external audit.

# Official 1.5 loop (2026-08-19)
propose harder tasks
generate / refine scaffold
produce solution rollouts
R_task = V * D(p*=0.2) * N
optimize three stages with GRPO

How to read the boundaries

“On par with Opus” ≠ ahead on every row

On the authors’ own DeepSWE column the 397B trails Opus 4.8 (56.0 vs 59.0). Terminal-Bench leads also depend on the harness (Terminus-2 vs Claude Code).

A self-made curriculum can still be hacked

The post defines a separate harness reward for task alignment, reward fidelity and hack resistance. Eval footnotes say SWE-bench dropped git history and cut the network, and NL2Repo blocked listed repos and pip. Those are their stated controls, not a proof the system cannot be gamed.

What is open is the checkpoint, not the full training bill

Weights, serving notes and quantized formats were published. The compute and data traces of the self-generated curriculum are not written as a line-by-line reproduction budget.

Questions worth checking