Model Details
Ornith-1.5-35B-A3B
Quick answer: Ornith-1.5-35B-A3B is the mid-size mixture-of-experts member of Ornith's (DeepReinforce's) Ornith-1.5 family — a 35B-total, ~3B-activated MoE model released under MIT license. It scores 79% on SWE-bench Verified, 67.8% on Terminal Bench 2.1, and 89.2% on GPQA Diamond, significantly outperforming its similarly-sized peer Qwen 3.6-35B and dense models like Gemma 4-31B and Meta's Muse Glimmer-30B despite activating only ~3B parameters per token.
At a Glance
Where Ornith-1.5-35B-A3B leads
- 79% SWE-bench Verified, 68.5% Terminal Bench 2.1 (Claude Code harness) — outperforms Gemma 4-31B (52.0% SWE-bench Verified) and Muse Glimmer-30B (76.0%) by wide margins on agentic coding despite activating only ~3B parameters
- 89.2% GPQA Diamond — strong graduate-level science reasoning for its size class
- 72.5% ClawEval, 67.6% BrowseComp — solid agentic tool-use and web-browsing performance
- MIT license — fully open weights, permissive for commercial use
- Efficient MoE design (~3B active) makes it deployable on 2× 80GB GPUs
Where it lags
- 22% DeepSWE — noticeably weaker than its flagship 397B sibling (56%) on this harder coding benchmark
- 33.4% HLE (with tools) / 25.6% (no tools) — modest on frontier-exam reasoning at this scale
- 5.1% Frontier-Bench v0.1 — reflects the benchmark's extreme difficulty across all but flagship-tier models
Best for: cost-efficient self-hosted coding agents and tool-use workflows where MoE efficiency (only ~3B active parameters) needs to significantly outperform similarly-sized dense models.
What Ornith-1.5-35B-A3B Is
Ornith-1.5-35B-A3B is the mid-size mixture-of-experts model in Ornith's (built by the DeepReinforce team) Ornith-1.5 family, sitting between the 9B dense edge model and the 397B flagship. Like its siblings, it inherits Ornith-1.5's self-improvement training loop — the model jointly generates its own training tasks, task-specific scaffolds, and solution rollouts, optimized end-to-end via reinforcement learning (GRPO), rather than relying on a fixed human-curated training distribution.
Despite activating only ~3B parameters per token, Ornith-1.5-35B-A3B "significantly outperforms its similar-sized peer Qwen 3.6-35B across all coding and agentic benchmarks," per the official model card, and beats larger dense models — Gemma 4-31B and Meta's Muse Glimmer-30B — by wide margins on agentic coding (68.5% vs. 43.4% and 51.7% on Terminal Bench 2.1's Claude Code harness; 79.0% vs. 52.0% and 76.0% on SWE-bench Verified). It supports a 262,144-token context window, extensible to roughly 1,000,000 tokens via YaRN RoPE scaling, and serves comfortably on 2× 80GB GPUs.
Specs
| Field | Value |
|---|---|
| Organization | Ornith (DeepReinforce) |
| Architecture | Mixture-of-Experts (Qwen3.5 MoE foundation) |
| Total parameters | 35B (36B per HF card) |
| Activated parameters | ~3B |
| Context window | 262,144 tokens (extensible to ~1M via YaRN) |
| Modality | Text (reasoning model — <think> traces by default) |
| License | MIT |
| Release date | August 2026 |
Pricing
Ornith-1.5-35B-A3B is released as open weights (self-hosted); Ornith/DeepReinforce has not announced a hosted API pricing tier for this model.
Public Benchmark Scores
| Benchmark | Score | Source | Date |
|---|---|---|---|
| Terminal Bench 2.1 | 67.8% | Ornith-1.5-35B-A3B model card (HF Eval Results, Terminus-2 harness) | 2026-08 |
| SWE-bench Verified | 79% | Ornith-1.5-35B-A3B model card (HF Eval Results) | 2026-08 |
| SWE-bench Pro | 59.6% | Ornith-1.5-35B-A3B model card (HF Eval Results) | 2026-08 |
| DeepSWE | 22% | Ornith-1.5-35B-A3B model card (HF Eval Results) | 2026-08 |
| GPQA Diamond | 89.2% | Ornith-1.5-35B-A3B model card (HF Eval Results) | 2026-08 |
| Humanity's Last Exam | 33.4% | Ornith-1.5-35B-A3B model card (HF Eval Results, with tools) | 2026-08 |
| MCP-Atlas | 70.2% | Ornith-1.5-35B-A3B model card (benchmark appendix) | 2026-08 |
| Toolathlon-Verified | 48.7% | Ornith-1.5-35B-A3B model card (benchmark appendix) | 2026-08 |
| BrowseComp | 67.6% | Ornith-1.5-35B-A3B model card (benchmark appendix) | 2026-08 |
Scores are self-reported by the Ornith team on the official Hugging Face model card (ornith-ai/Ornith-1.5-35B-A3B) and the accompanying technical blog post; the top 6 rows are additionally confirmed via the card's structured HF "Evaluation Results" metadata. All results are averaged over 5 independent runs. HLE score uses the "with tools" evaluation condition (25.6% without tools).
Notable Results Not Yet Tracked as Benchgen Benchmarks
- SWE-bench Multilingual (71.4%), Frontier-Bench v0.1 (5.1%), NL2Repo (46.2%), SWE Atlas – QnA (39.8%), WideSearch (67.8%), and ClawEval (72.5%) 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 (same set flagged on the sibling Ornith-1.5-397B page).
Ornith-1.5-35B-A3B vs Alternatives
| Model | Terminal Bench 2.1 (Claude Code) | SWE-bench Verified | Active Params |
|---|---|---|---|
| Ornith-1.5-35B-A3B | 68.5% | 79.0% | ~3B |
| Gemma 4-31B | 43.4% | 52.0% | 31B (dense) |
| Muse Glimmer-30B | 51.7% | 76.0% | 30B (dense) |
Despite activating far fewer parameters than either dense competitor, Ornith-1.5-35B-A3B outperforms both on agentic coding benchmarks — a notable efficiency result for its MoE design.
FAQ
Is Ornith-1.5-35B-A3B open source?
Yes — MIT license, with weights available on Hugging Face (ornith-ai/Ornith-1.5-35B-A3B), plus FP8, NVFP4, and GGUF quantized variants.
Who makes Ornith-1.5-35B-A3B? Ornith, built by the DeepReinforce team.
How big is Ornith-1.5-35B-A3B? 35B total parameters (Mixture-of-Experts), with only ~3B activated per forward pass.
What is Ornith-1.5-35B-A3B's context window? 262,144 tokens natively, extensible to roughly 1,000,000 tokens via YaRN RoPE scaling.
How does Ornith-1.5-35B-A3B compare to similarly-sized models? It significantly outperforms Qwen 3.6-35B across coding and agentic benchmarks, and beats larger dense models (Gemma 4-31B, Muse Glimmer-30B) on agentic coding despite activating far fewer parameters per token.
Where can I access Ornith-1.5-35B-A3B? Self-hosted from the open weights at huggingface.co/ornith-ai/Ornith-1.5-35B-A3B (BF16, FP8, or NVFP4), via vLLM or SGLang on 2× 80GB GPUs.
Benchmark scores sourced from the official Hugging Face model card for ornith-ai/Ornith-1.5-35B-A3B and the Ornith team's technical blog (Aug 2026), cross-checked against the card's structured HF Eval Results metadata where available.