← Back to feed

DeepReinforce has released Ornith-1.5, a three-size open model family (397B, 35B, 9B) that extends the self-scaffolding framework introduced in Ornith-1.0 into a full self-improvement loop. The key architectural change: Ornith-1.5 proposes its own training tasks rather than learning from a fixed set of human-curated problems.

What changed from 1.0

Ornith-1.0 wrote the scaffolding code around a human-defined task list. The model improved the process of solving problems humans selected, but the problems themselves were static. The ceiling was whatever the curated set covered.

Ornith-1.5 removes that constraint. The model proposes tasks, generates training data from those proposed tasks, and trains on the result. The loop is: propose, generate, train, repeat — with no human selection step in the middle. The practical effect is that Ornith-1.5 can target capability gaps that no human curator explicitly identified.

Three sizes, one loop

The release spans:

• 397B — the flagship, same parameter count as Ornith-1.0’s largest variant
• 35B — mid-range, positioned as the deployable option for teams running their own infrastructure
• 9B — the edge-capable version; Ornith-1.0’s 9B scored 69.4% on SWE-bench Verified

The self-improvement loop applies across all three sizes. Smaller models benefit proportionally more when the task generation is calibrated to their current capability level rather than to a frontier-level task distribution.

What DeepReinforce is claiming

The framing from the Ornith-1.5 announcement is that this is an “end-to-end self-improvement loop” — a distinction from prior work that either kept humans in the task selection step or required a stronger model to generate training data for a weaker one. Ornith-1.5 does neither: the same model that will be trained is also the model proposing and generating the tasks.

That claim requires scrutiny. Self-generated tasks can produce degenerate training distributions — tasks the model already knows how to solve, creating the appearance of improvement without genuine capability gain. The published materials from DeepReinforce do not yet include full benchmark comparisons against Ornith-1.0 on held-out evaluations, which is the standard that would validate the loop is generating real generalization rather than overfitting to self-selected easy cases.

Open weights

Ornith-1.0 shipped under the MIT license, which made it one of the more permissively licensed frontier-adjacent coding models available. The 1.5 release continues that pattern. The 397B variant’s open weights place it alongside Kimi K3 (2.8T MoE, also open weight) as one of the few genuinely open models in the high-SWE-bench tier.

Benchmark results for Ornith-1.5 on SWE-bench Verified and Terminal-Bench are pending independent confirmation.