Ornith-1.5-35B-A3B · REAP-50 (pruned bf16)

The 50%-REAP-pruned Ornith-1.5-35B-A3B in bf16 — the source checkpoint behind the NVFP4 and GGUF releases. Provided so others can produce their own quants (AWQ, EXL2, MLX, custom GGUF, …).

• 256 → 128 experts via REAP (Router-weighted Expert Activation Pruning), with a router-renormalization fix over the survivors.
• MTP head removed (mtp_num_hidden_layers: 0), reload-verified with a real forward pass.
• ~19B parameters, qwen3_5_moe hybrid (Gated-DeltaNet + attention + MoE).
• Vision tower not stripped in this checkpoint — use Qwen3_5MoeForCausalLM for text-only, or see the pipeline for the vision-strip step.

Quality after NVFP4A16 quant: HumanEval+ 84.2%, MBPP+ 89.2% (greedy, instruct) — single greedy draws; see the NVFP4A16 card on run-to-run spread.

Architecture

graph TD
    Base["ornith-ai/Ornith-1.5-35B-A3B<br/>qwen3_5_moe - 256 experts - MTP head - ~65 GB bf16 - MIT"]
    subgraph Build ["Build - RTX 5070 Ti, SM120"]
        REAP["REAP expert prune 50%  256 -> 128"]
        MTP["strip MTP head (1 -> 0, reload-verified)"]
        StripV["strip vision tower (333 tensors)"]
        Quant["GPTQ-NVFP4A16 quantize (~5.5 h, 12.47 GiB)"]
    end
    subgraph HF ["Published formats"]
        A16["REAP-50-NVFP4A16 - 12.47 GiB (vLLM)"]
        GGUF["REAP-50-GGUF (Q4_K_M / Q5_K_M / Q6_K / Q8_0)"]
        BF16["REAP-50-bf16 (pruned source)"]
    end
    Bench["HumanEval+ 84.2% - MBPP+ 89.2% - SWE-bench Verified 44.0%"]
    Base --> REAP --> MTP --> StripV --> Quant --> A16
    StripV --> BF16
    BF16 -. convert .-> GGUF
    A16 --> Bench

Releases built from this

• NVFP4A16 (vLLM / Blackwell): Ttimms/Ornith-1.5-35B-A3B-REAP-50-NVFP4A16
• GGUF: Ttimms/Ornith-1.5-35B-A3B-REAP-50-GGUF

Pipeline

Full prune → strip → quant → serve → evaluate pipeline, with the bugs found and fixed along the way: https://github.com/t-timms/ornith-nvfp4

License

MIT (inherits from ornith-ai/Ornith-1.5-35B-A3B). Pruning via REAP (github.com/CerebrasResearch/reap).