An open-weights model family released this week claims agentic coding parity with Anthropic's strongest commercial system, and it arrives in a size that fits on a phone. DeepReinforce's Ornith-1.5 ships in three tiers — a 397B mixture-of-experts flagship, a 35B MoE that activates only 3B parameters per token, and a 9B dense model with a quantized Mobile build for iPhone and Android.
On Terminal-Bench 2.1, averaged over five independent runs, the flagship records 86.1 against the 85.0 the team attributes to Claude Opus 4.8, while trailing on DeepSWE at 56.0 to 59.0. The comparison that matters more for the open ecosystem is closer to home: GLM-5.2 and DeepSeek-V4-Flash-0731 both stall at 82.7 on Terminal-Bench, with DeepSWE results of 46.2 and 54.4. Flagship coverage extends past code into 92.8 on GPQA Diamond and 86.6 on BrowseComp.
A Model That Writes Its Own Curriculum
The interesting claim here is not the leaderboard position but the provenance of the training data. Ornith-1.0, released in June, made the scaffold — the instructions, tools, decomposition strategy and orchestration wrapped around a model — a learnable object rather than something engineers hand-build. Version 1.5 extends that to the task list itself, so curriculum, evaluation harness and answers all fall out of the same optimization run.
Each cycle proposes problems harder than what the model has already solved, generates a task-specific scaffold for each, and produces a solution rollout conditioned on both, with reward propagating back through all three stages under GRPO. Task reward multiplies validity, frontier difficulty and novelty, and validity works as a hard gate that zeroes the entire term whenever a generated task or harness fails to form a verifiable environment. Difficulty is measured from the model's own rollout success rate against a 0.2 target, which quietly guarantees that a problem stops paying the generator as soon as the policy learns to clear it.
That last detail is the closest thing to a novel idea in the release. Most synthetic-data pipelines decay because the generator drifts toward whatever it can already produce cleanly; pegging reward to a low success rate inverts that incentive and forces the curriculum to stay uncomfortable. Whether it holds across many more training cycles than the ones reported is the open question, and it is not one a benchmark table can answer.
The Smaller Tiers Carry the Practical Story
The 35B reaches 68.5 on Terminal-Bench 2.1 and 79.0 on SWE-Bench Verified on a 3B active-parameter budget, against 43.4 and 52.0 for Gemma 4-31B and 51.7 for Meta's Muse Glimmer-30B on the agentic benchmark. The 9B lands at 47.0 and 70.6, above both Gemma 4-31B and Qwen 3.6-35B despite the parameter gap.
For most teams that mid tier, not the flagship, is the release. A 397B MoE is a datacenter commitment with a serving bill attached; a 35B that activates 3B per token is something a single well-specified node can host, and it is the first open checkpoint at that size to post agentic coding numbers in the same neighborhood as models several times larger. If the figures survive outside evaluation, the practical effect is downward pressure on inference pricing for coding agents, not a leaderboard reshuffle.
Reading the Numbers Carefully
Three caveats deserve weight. These are vendor-published results, and the evaluation leans on rival infrastructure — Claude 4.6 Opus and Claude 4.8 Opus act as judge models on HLE and MCP-Atlas, so a competitor's system is scoring the work. The lineage matters too, since Ornith-1.0 was post-trained on Qwen 3.5 and Gemma 4 checkpoints, meaning the self-improvement loop compounds gains on pretraining someone else financed. And an 86.1 that sits 1.1 points above a rival is well inside the range where harness configuration and prompt formatting can decide the ordering.
The anti-hacking discipline is more reassuring than the industry norm. Git history was stripped from repository images to block retrieval of prior fixes, network access was disabled during SWE-Bench runs, and repository and package access was cut off on NL2Repo. For a system whose entire training signal is self-generated, that plumbing is the part independent evaluators should probe first.
