Ornith-1.5-397B

// Overview

Ornith-1.5-397B is the flagship of the Ornith-1.5 family, released on 18 August 2026 by Ornith — the model lab of the DeepReinforce team — alongside a 35B Mixture-of-Experts sibling and a 9B dense one. It is a 397-billion-parameter Mixture-of-Experts model that emits a thinking block before its final answer, and it ships as downloadable weights under the MIT license.

What distinguishes the family is how it is trained. Ornith-1.0, built on top of open Qwen3.5 and Gemma 4 bases with additional continued pretraining, mid-training and post-training, already optimised the scaffold and the rollouts around a fixed set of human-curated tasks. Ornith-1.5 extends that loop end to end: the model proposes new training tasks that expose its own capability gaps, constructs a task-specific scaffold of instructions, tools and decomposition strategies for each, and then produces the solution rollouts used for reinforcement learning — with reward propagated across all three stages. Ornith describes the result as a closed self-improvement loop in which a stronger policy generates harder and more informative tasks.

The model uses a 60-layer architecture with 512 experts and 10 routed per token, interleaving linear-attention layers with a full-attention layer every fourth layer. Its native context is 262,144 tokens and Ornith documents YaRN RoPE scaling at a factor of 4.0 to reach roughly one million tokens for workloads that need it. In bf16 the checkpoint is about 800 GB, so Ornith's own quickstart assumes multi-GPU serving with vLLM ≥ 0.19.1 or SGLang ≥ 0.5.9 at tensor-parallel size 8; FP8, NVFP4 and GGUF builds are published for smaller deployments, and the GGUF build runs under llama.cpp and Ollama.

On Ornith's published evaluations the 397B model scores 86.1 on Terminal-Bench 2.1 and 86.0 on SWE-bench Verified, ahead of Claude Opus 4.8 on both (85.0 and 85.8) and ahead of the open-weight GLM-5.2 and DeepSeek-V4-Flash-0731. The picture is not uniform: on the harder agentic rows — Frontier-Bench v0.1, NL2Repo, DeepSWE and SWE Atlas QnA — it trails Opus 4.8 by a wide margin, and Kimi K3 leads it on Terminal-Bench, DeepSWE and BrowseComp. It leads every listed model on WideSearch (80.8) and BrowseComp among the open field.

// Benchmarks

This model's scores

1. GPQA Diamond92.8%
2. BrowseComp86.6%
3. Terminal-Bench 2.1 (Terminus-2)86.1%
4. SWE-bench Verified86%
5. ClawEval81.4%
6. WideSearch80.8%
7. MCP-Atlas80%
8. SWE-bench Multilingual79.6%
9. Toolathlon-Verified71.2%
10. SWE-bench Pro65.1%
11. NL2Repo59.5%
12. DeepSWE56%
13. HLE (with tools)56.1%
14. HLE (no tools)44.6%
15. Frontier-Bench v0.113.5%

Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.

// Strengths

• MIT-licensed open weights, free to download, fine-tune, self-host and use commercially
• Terminal-Bench 2.1 86.1 and SWE-bench Verified 86.0 on Ornith's own evaluations — ahead of Claude Opus 4.8 on both rows
• 262K native context, documented to extend to roughly 1M tokens with YaRN scaling at factor 4.0
• Sparse Mixture-of-Experts routing (10 of 512 experts per token) keeps per-token compute far below the 397B total
• First-class agent tooling: XML tool-call and reasoning parsers for vLLM and SGLang, plus published FP8, NVFP4 and GGUF builds

// Best for

• Long-horizon terminal and repository coding agents run on self-hosted infrastructure
• Deep-research and web-browsing agents (Ornith reports the family's best WideSearch and BrowseComp scores here)
• Teams that need frontier-adjacent agentic quality under a permissive license rather than a vendor API
• Research on self-improving training loops, where the released weights are the reference implementation of the approach

// Ornith (open-weight) — every version

The full lineage of the Ornith (open-weight) line, newest first. Every version has its own page — click any to compare specs, benchmarks and pricing.