Ornith-1.5-397B

Ornith-1.5-397B

Quick answer: Ornith-1.5-397B is the flagship member of Ornith's (DeepReinforce's) Ornith-1.5 family — a 397B-total mixture-of-experts model trained with a self-improvement loop that jointly generates its own training tasks, scaffolds, and solution rollouts. Released under MIT license, it scores 86.1% on Terminal Bench 2.1, 86% on SWE-bench Verified, and 92.8% on GPQA Diamond — performing on par with Claude Opus 4.8 while outperforming similarly-sized open models like GLM-5.2 and DeepSeek-V4-Flash-0731.

At a Glance

Where Ornith-1.5-397B leads

• 86.1% Terminal Bench 2.1 (Terminus-2 harness) — on par with Claude Opus 4.8 (85.0%)
• 86% SWE-bench Verified — strong real-world coding agent performance
• 92.8% GPQA Diamond — near-flagship-tier graduate-level science reasoning
• 86.6% BrowseComp — strong web-browsing agentic search
• 56% DeepSWE — competitive with closed flagship models on this harder coding benchmark
• MIT license — fully open weights, permissive for commercial use

Where it lags

• 65.1% SWE-bench Pro — notably behind SWE-bench Verified, consistent with SWE-bench Pro's harder task design across all models
• 13.5% Frontier-Bench v0.1 — reflects the benchmark's extreme difficulty rather than a specific weakness
• 56.1% HLE (with tools) / 44.6% (no tools) — trails top closed frontier models on the hardest expert-knowledge questions

Best for: self-hosted flagship-scale coding agents and long-horizon agentic workflows where MIT licensing and full weight access matter, and where near-Claude-Opus-4.8-tier performance is needed without a closed API.

What Ornith-1.5-397B Is

Ornith-1.5-397B is the largest and most capable member of Ornith's (built by the DeepReinforce team) newly released Ornith-1.5 model family, which spans three scales: this 397B mixture-of-experts flagship, a 35B-A3B mid-size MoE, and a 9B dense edge model. Ornith-1.5 builds on Ornith-1.0 (itself developed on top of Qwen3.5 and Gemma 4 with continued pretraining, mid-training, and post-training), extending Ornith-1.0's "self-scaffolding" approach into a more complete self-improvement loop.

Instead of training on a fixed set of human-curated tasks with manually designed evaluation harnesses, Ornith-1.5 continuously generates its own training tasks, discovers effective scaffolds (instructions, tools, decomposition strategy) for solving them, and produces solution rollouts — all optimized jointly via reinforcement learning (GRPO) using validity, frontier-difficulty, and novelty reward signals. This closed loop lets the model expand its own curriculum as it improves, rather than being bottlenecked by a static training distribution.

Ornith-1.5-397B natively supports a 262,144-token context window, extensible to roughly 1,000,000 tokens via YaRN RoPE scaling (factor 4.0). Per the official benchmark results, it scores 86.1% on Terminal Bench 2.1 and 56% on DeepSWE — performing on par with Claude Opus 4.8 (85.0% / 59.0%) and outperforming leading open-source models of similar scale, including GLM-5.2 (81% / 46.2%) and DeepSeek-V4-Flash-0731 (82.7% / 54.4%).

Specs

Pricing

Ornith-1.5-397B is released as open weights (self-hosted); Ornith/DeepReinforce has not announced a hosted API pricing tier for this model.

Public Benchmark Scores

Scores are self-reported by the Ornith team on the official Hugging Face model card (ornith-ai/Ornith-1.5-397B) and the accompanying technical blog post; the top 6 rows are additionally confirmed via the card's structured HF "Evaluation Results" metadata (the most reliable source, unambiguously tying each score to this specific model). All results are averaged over 5 independent runs per the model card's methodology notes. HLE score uses the "with tools" evaluation condition (44.6% without tools) — see the "no tools" caveat in Benchgen's Humanity's Last Exam page.

Notable Results Not Yet Tracked as Benchgen Benchmarks

Several benchmarks reported on the model card don't yet have standalone Benchgen pages:

• SWE-bench Multilingual (79.6%, HF dataset SWE-bench/SWE-bench_Multilingual), Frontier-Bench v0.1 (13.5%), NL2Repo (59.5%), SWE Atlas – QnA (55.6%), WideSearch (80.8%), and ClawEval (81.4%) are all real, methodology-documented benchmarks per the model card's footnotes, but don't have existing Benchgen benchmark pages yet — flagged as follow-ups.

Ornith-1.5-397B vs Alternatives

Per the official model card, Ornith-1.5-397B performs on par with Claude Opus 4.8 on both Terminal Bench 2.1 and DeepSWE while outperforming similarly-sized open-weight peers — a notable result for an MIT-licensed, fully open model.

FAQ

Is Ornith-1.5-397B open source? Yes — MIT license, with weights available on Hugging Face (ornith-ai/Ornith-1.5-397B), plus FP8, NVFP4, and GGUF quantized variants.

Who makes Ornith-1.5-397B? Ornith, built by the DeepReinforce team.

How big is Ornith-1.5-397B? 397B total parameters (Mixture-of-Experts), 16B activated per forward pass.

What is Ornith-1.5-397B's context window? 262,144 tokens natively, extensible to roughly 1,000,000 tokens via YaRN RoPE scaling.

How does Ornith-1.5-397B compare to closed frontier models? It performs on par with Claude Opus 4.8 on Terminal Bench 2.1 (86.1% vs 85.0%) and DeepSWE (56% vs 59%), while being fully open-weight under MIT.

Where can I access Ornith-1.5-397B? Self-hosted from the open weights at huggingface.co/ornith-ai/Ornith-1.5-397B (BF16, FP8, or NVFP4), via vLLM or SGLang on an 8-GPU node.

Benchmark scores sourced from the official Hugging Face model card for ornith-ai/Ornith-1.5-397B and the Ornith team's technical blog (Aug 2026), cross-checked against the card's structured HF Eval Results metadata where available.