SYLIN

Ornith

by Ornith AI · Model · Language & code models

Ornith 1.5 open family: 9B dense, 35B MoE and 397B MoE, trained with a self-improvement loop generating tasks, verifiers and RL data

AccessOpen weightsNeeds~18.8 GBDownload18.8 GBLicenseMIT

Versions

VersionReleasedRank
Ornith 1.5LATESTAug 19
Ornith 1.0

Evidence

  1. Aug 28
    Claimed to beat similarly sized Qwen 3.6 35B-A3B on every coding and agentic benchmark, and some larger dense models
  2. Aug 28
    Given one open-ended goal, it planned, built, tested and verified a real-time price alert feature in a three-container crypto dashboard
  3. Aug 23
    397B claimed to beat GLM 5.2 (twice its size) and come close to closed Opus 4.8 on agentic coding and knowledge-work benchmarks
  4. Aug 20
    Reviewer found the 35B MoE variant quite good in earlier tests, far better than the 9B
  5. Aug 19
    Ornith-1.5 released in three sizes: 397B MoE, 35B-A3B MoE and 9B dense, plus a 9B-Mobile build for phones
Show all 8
  1. Aug 19
    Model proposes harder tasks, writes its own scaffolds and solves them, all three stages trained jointly with GRPO
  2. Aug 19
    397B scores 86.1 on Terminal-Bench 2.1 versus 85.0 for Claude Opus 4.8 and 82.7 for DeepSeek-V4-Flash; trails Opus on DeepSWE (56.0 vs 59.0)
  3. Aug 19
    Unlike the explicitly non-general-purpose Ornith 1, version 1.5 targets general reasoning, agentic and coding use

Cost to run

Alternatives

Ornith 1.5 9BStruggled on a single-file HTML rotisserie chicken coding task, producing garbage and hallucinating in Chinese; reviewer calls it unusable
More Ornith
Made from