← Back to feed

DeepReinforce has released Ornith-1.0, a family of MIT-licensed agentic coding models spanning four sizes: 9B dense, 31B dense, 35B MoE, and the flagship 397B MoE with 17B active parameters. The lab’s self-reported benchmark results put the 397B variant at 82.4% on SWE-Bench Verified and 77.5% on Terminal-Bench 2.1 — ahead of every previously published open-weight result on both leaderboards.

The models are post-trained on Gemma 4 and Qwen 3.5 base checkpoints.

The Training Method

The differentiating claim in Ornith is its training objective. Standard RL-based coding agents learn to produce better solutions, optimizing for whether the final answer is correct. Ornith’s training jointly optimizes the scaffold — the plan, memory pattern, tool selection sequence, error handling approach, and search strategy — alongside the solution rollout.

In the training loop, the model first proposes a scaffold for a given task, then uses that scaffold to generate solution rollouts. Reward signals propagate through both stages simultaneously. The effect, according to DeepReinforce’s tech blog, is that the model learns which scaffolds work for which categories of task rather than defaulting to a single rigid checklist across all problem types.

This mirrors the intuition behind research like SWE-RL and Meta’s self-play coding work, but extends it from solution quality to scaffold quality as a co-optimized target.

The Numbers

The 397B MoE results on DeepReinforce’s in-house evaluation:

Benchmark	Score
SWE-Bench Verified	82.4%
Terminal-Bench 2.1	77.5%
SWE-Bench Multilingual	78.9%
NL2Repo	48.2
SWE Atlas QnA	41.2
ClawEval	77.1

The 9B dense model posts 69.4% on SWE-Bench Verified — a notable result for a model that fits on consumer hardware.

These numbers are self-reported by the lab and have not yet been independently replicated on the standard SWE-bench.com leaderboard. The comparison to frontier proprietary models should be read with that caveat: the gap between vendor-reported and third-party-evaluated scores can be material, as seen repeatedly across the 2026 benchmark landscape.

Context on the Open-Source Coding Race

The open-source coding tier has moved fast. Alibaba’s Qwen3.6-35B-A3B scored 73.4% on SWE-Bench Verified in April. DeepSeek V4 Flash reached 79.0%. Kimi K2.7-Code claimed 81% with verified MCP results. Ornith’s 82.4% claim, if it holds under independent evaluation, would extend that progression and narrow the gap with proprietary frontier models whose published SWE-Bench Verified scores cluster in the 80-88% range (excluding Claude Fable 5 and Mythos, which are in a different tier).

The MIT license matters for commercial deployment. Apache 2.0 and MIT-licensed frontier-class models remain rare at the top of the coding benchmark stack.

Practical Use

With 17B active parameters on a 397B MoE, the flagship Ornith variant runs at roughly the cost of a 17B dense model per forward pass, similar to DeepSeek’s MoE architecture approach. That makes it economically competitive with Qwen3 and Kimi K2.7 for hosted inference, and deployable on multi-GPU rigs for on-premises use. Model weights are available on HuggingFace. DeepReinforce’s technical blog at deep-reinforce.com documents the scaffold optimization methodology.