Ornith-1.0-35B — PrismaAURA 4.75-bit (mixed NVFP4/FP8/BF16) + MTP
Mixed-precision quantization of deepreinforce-ai/Ornith-1.0-35B
via PrismaAURA — per-Linear bit allocation selected on real end-to-end KL, shipped as a stock
compressed-tensors checkpoint that vanilla vLLM serves with no forked runtime and no custom kernels.
Includes a Multi-Token-Prediction (MTP) head for speculative decoding.
| Metric | Value |
|---|---|
| Effective rate (body) | 4.75 bits / quantizable parameter |
| Size | ~23 GB (from ~70 GB BF16) |
| Served confident KL-vs-BF16 | 0.0143 (top-1 agreement 98.6%) |
| MTP acceptance | 91.3% @ pos-0, 77.3% overall (~2.32 accepted / 3 drafted) |
| Serving | vanilla vLLM, CUTLASS native NVFP4 W4A4 on Blackwell |
MTP note
The base Ornith checkpoint ships no MTP weights (the fine-tune stripped them). This artifact grafts the base Qwen3.5-35B-A3B MTP head (BF16), the same head every community MTP-Ornith variant uses. Because it predicts the base distribution rather than Ornith's RL-tuned one, acceptance is a bit lower than a native head would give — but still strong (91% at position 0).
Serving
With speculative decoding (MTP):
vllm serve rdtand/Ornith-1.0-35B-PrismaAURA-4.75bit-vllm-MTP \
--trust-remote-code --attention-backend flashinfer --moe-backend marlin \
--gpu-memory-utilization 0.85 --max-model-len 32768 --max-num-batched-tokens 8192 \
--enable-chunked-prefill \
--speculative-config '{"method":"mtp","num_speculative_tokens":3}'
Without MTP (e.g. for clean perplexity measurement — spec-decode poisons logprob metrics), drop the
--speculative-config flag. --max-num-batched-tokens >= 2096 is required (Gated-DeltaNet page size).
License
Inherits the base model's license — see deepreinforce-ai/Ornith-1.0-35B.
Contact
Robert Tand — robert.tand@icloud.com