Ornith-1.5-35B-A3B-REAP-20pct

Ornith-1.5 coding MoE with 20% of experts gone. Same 8 active per token, ~56 GB BF16 instead of ~67 GB. REAP (Lasby et al., Cerebras, ICLR 2026). Not lossless. Not GGUF. Not a 24 GB card.

Same recipe on Qwen3.5-35B: 0xSero/Qwen3.5-28B.

Model details

Shared expert, vision tower, embeddings, and lm_head are untouched. Router weights renormalized after prune. config.json text_config.num_experts is 205. Architecture: Qwen3_5MoeForConditionalGeneration.

Quickstart

pip install -U "transformers>=5.8.1"

from transformers import AutoModelForImageTextToText, AutoTokenizer

model_id = "oxfrug/Ornith-1.5-35B-A3B-REAP-20pct"
tok = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForImageTextToText.from_pretrained(
    model_id, torch_dtype="auto", device_map="auto"
)

messages = [{"role": "user", "content": "Write a one-line Python lambda that squares a number."}]
text = tok.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tok(text, return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=256, temperature=0.6, top_p=0.95)
print(tok.decode(out[0][inputs.input_ids.shape[1]:], skip_special_tokens=True))

vLLM (>=0.19.1), 2× 80GB recommended for long context:

vllm serve oxfrug/Ornith-1.5-35B-A3B-REAP-20pct \
    --tensor-parallel-size 2 \
    --max-model-len 262144 \
    --enable-auto-tool-choice --tool-call-parser qwen3_xml \
    --reasoning-parser qwen3 \
    --trust-remote-code

Sampling: general temperature=0.6, top_p=0.95, top_k=20. Ornith is a reasoning model — the assistant turn opens with <think>…</think> unless the server splits it into reasoning_content.

Benchmarks

This checkpoint was not re-run on SWE-bench / Terminal-Bench / GPQA. Do not copy the base card’s 79 SWE-bench / 67.8 Terminal-Bench numbers onto this prune.

vs base (this checkpoint)

Proxy (same recipe, not this checkpoint)

The same REAP 20% recipe on Qwen/Qwen3.5-35B-A3B (0xSero/Qwen3.5-28B, greedy, 0-shot) dropped coding a few points and raised WikiText perplexity. Ornith is a Qwen3.5 hybrid (GatedDeltaNet + MoE); expect a comparable tax, not these exact numbers.

Source: 0xSero’s card, not a run of oxfrug/Ornith-1.5-35B-A3B-REAP-20pct.

Smoke (this checkpoint)

Greedy, thinking off, 3090 offload. Not a published coding bench.

How it was made

Same mix as 0xSero/Qwen3.5-28B, not Cerebras’ 24k×16k agentic mix:

• theblackcat102/evol-codealpaca-v1 × 250
• open-r1/Mixture-of-Thoughts code / math / science × 250 each
• 1000 sequences, max length 1024 (3090 VRAM; 0xSero used 4096)
• seed 42, prune_method=reap, renormalize_router_weights=true
• Layerwise observer (model does not fit in 60 G RAM)

No xlam / SWE-smith traces — extra risk for Ornith’s tool-calling specialists.

Intended use

Local / self-hosted coding and tool-calling with Ornith-1.5 quality, at ~11 GB less disk than the unpruned BF16. Accept a quality tax vs the 35B-A3B base.

Limits

• Expert prune, not a quant and not a fine-tune.
• Shorter calibration sequences (1024 vs 4096) than the Qwen3.5-35B sibling run.
• Hybrid GatedDeltaNet + MoE: REAP only touches MoE experts.
• Do not claim base Ornith-1.5 SWE-bench / Terminal-Bench / GPQA scores for this file.

License

MIT, same as the base. Prune by oxfrug.

Citation

@misc{lasby2025reap,
  title  = {REAP the Experts: Why Pruning Prevails for One-Shot MoE Compression},
  author = {Mike Lasby and Ivan Lazarevich and Nish Sinnadurai and Sean Lie and Yani Ioannou and Vithursan Thangarasa},
  year   = {2025},
  eprint = {2510.13999},
  archivePrefix = {arXiv}
}

@misc{ornith_1_5,
  title  = {{Ornith-1.5}: From Self-Scaffolding to Self-Improvement},
  url    = {https://ornith.ai/ornith_1_5.html},
  author = {{Ornith Team}},
  year   = {2026}
}