Ornith 1.5 35B A3B

One-line: Self-improving 3B-active MoE that beats every same-size peer and many larger dense models on agentic coding (SWE-bench Verified 79, MCP-Atlas 70.2) — the strongest genuinely local-runnable coding agent in this sweep.

Coding benchmarks

Benchmark	Score	Source
HumanEval	—	—
HumanEval+	—	—
MBPP	—	—
LiveCodeBench	—	—
Aider polyglot	—	—
SWE-bench Verified	79	lab (OpenHands harness, lab-reported)
SWE-bench Multilingual	71.4	lab (OpenHands harness, lab-reported)
SWE-bench Pro	59.6	lab (OpenHands harness, lab-reported)
Terminal-Bench 2.1	67.8	lab (Terminus-2 framework, lab-reported)
DeepSWE	22	lab (Claude Code harness, lab-reported)
NL2Repo	46.2	lab (lab-reported)

Agentic / tool-use benchmarks

Benchmark	Score	Source
BFCL	—	—
τ-Bench	—	—
ToolACE	—	—
MCP-Atlas	70.2	lab (lab-reported)
Toolathlon-Verified	48.7	lab (official eval service, lab-reported)
ClawEval	72.5	lab (lab-reported)

Reasoning (context only)

Benchmark	Score	Source
MMLU / MMLU-Pro	—	—
GPQA Diamond	89.2	lab
MATH	—	—
AIME 2025	—	—
HLE (no tools)	25.6	lab
HLE (with tools)	33.4	lab

Efficiency

• Active params: ~3B active per token (35B total; 256 experts, top-8 routed + shared)
• Context length: 262,144 tokens (256k); YaRN ×4 → ~1M
• VRAM (fp16 rough): ~70 GB in bf16 for full 256k context — README recommends 2×80GB GPUs; GGUF build available for llama.cpp / Ollama
• Architecture: MoE on Qwen3.5-MoE base with hybrid linear-attention layers (global full attention every 4 layers), bf16
• Runtime: Requires Transformers ≥ 5.8.1, vLLM ≥ 0.19.1, SGLang ≥ 0.5.9; emits <think> reasoning + <tool_call> blocks for OpenAI-style tool_calls

What makes it notable

Ornith-1.5-35B-A3B is the strongest coding/agentic result at the 3B-active MoE scale in this sweep. It significantly outruns its direct peer Qwen3.6-35B-A3B (SWE-bench Verified 79 vs 73.4; Terminal-Bench 2.1 67.8 vs 52.5) and beats dense models like Gemma 4-31B and Muse Glimmer-30B on agentic coding by wide margins. The claim to fame is end-to-end self-improvement: the lab joint-optimizes task generation, harness construction, and solution rollouts via RL rather than fixed human-curated tasks — an extension of the Ornith-1.0 self-scaffolding loop. Tool-calling is first-class (MCP-Atlas 70.2, Toolathlon 48.7) and it ships OpenAI-compatible serving recipes for vLLM/SGLang with a tool-call parser. ⚠ All benchmark numbers are lab-reported (5-run averages) — no independent third-party confirmation found yet. Note its base (Qwen3.5-MoE) carries a vision config, but published evals are text/agentic-only.

See also

• welcome
• Source: https://huggingface.co/ornith-ai/Ornith-1.5-35B-A3B
• Blog: https://deep-reinforce.com/ornith.html