Wait Which Model?
← Back to directory
DeepReinforce

Ornith-1.5-397B

Released Aug 19, 2026 · knowledge cutoff unpublished

Status
Superseded
Location
United States
Type
LLM
Modality
Text
Context window
262K
Max output
—
Speed
—
Price ($/MTok in / out)
— / —
Cost per task
—
Open weights
Yes
Benchmarks5 of 10 reported

GPQA Diamond

92.8%

SWE-bench Verified

86%

SWE-bench Pro

65.1%

Terminal-Bench 2.1

86.1%

Humanity's Last Exam

44.6%

MMLU-Proretired

—

AIME (latest)retired

—

LMArena Elo

—

GDPval-AA v2

—

ARC-AGI-2

—
Strengths & weaknesses

Strengths

  • Trained with an end-to-end self-improvement loop that generates its own tasks, scaffolds and solution rollouts rather than inheriting a human-authored training harness
  • Early agentic tests on the family show it planning, building and self-verifying open-ended goals unassisted — including running its own test suite to confirm a fix before stopping
  • Ships all three sizes (9B, 35B-A3B, 397B) under a permissive MIT license with no gated weights

Weaknesses

  • Launch-week reception is thin and size-dependent — hands-on tests of the smaller 9B sibling exposed a wide gap between published benchmarks and real-world reliability, so the flagship's practical behavior is still largely unverified
  • Reported to perform poorly in plain chat without tool access, consistent with a model post-trained specifically for agentic scaffolds rather than general use
  • No first-party hosted API at launch — running the 397B checkpoint means self-hosting roughly 800GB (bf16) across multiple GPUs

SWE-bench Verified (86) and Terminal-Bench 2.1 (86.1) are the 397B variant's launch comparisons against Claude Opus 4.8 (85.8 / 85), widely repeated at launch; GPQA Diamond and HLE were initially left null because secondary sources conflicted; the HF model card has since supplied them (below). No official Artificial Analysis measurement or first-party API pricing published as of release. DeepReinforce's Ornith-1.0 (this model's obvious lineage) isn't tracked on this site, so predecessorId is left null rather than pointing at an untracked id. SWE-bench Pro 65.1, GPQA Diamond 92.8 and HLE 44.6 (no tools; 56.1 with tools) are Ornith AI's own model-card figures (huggingface.co/ornith-ai/Ornith-1.5-397B, averaged over five runs), which also give SWE-bench Multilingual 79.6 and DeepSWE 56; the earlier conflicting secondary GPQA/HLE figures are superseded by the card.

For developers

API model strings

Not researched

Licence

MIT License · Permissive · commercial use permitted

Retirement

No retirement announced

Lineage

No recorded predecessor

News