We let Claude Fable 5 perform brain surgery on a local model

We run Ornith 1.0 35B, a mixture of experts model, on a single secondhand Tesla P40 in the studio. It is a lovely model and a patient card, but generation hovered around 56 tokens per second. So we handed the problem to Claude Fable 5 with a one line brief: make this faster.

What it found was better than an optimization. Ornith is built on top of Qwen 3.6, same architecture, identical tensor dimensions, same 248,320 token vocabulary. Ornith has no published MTP head. Qwen3.6-35B-A3B-MTP does. If the shapes match, the head should transplant.

A quick note on MTP, or multi token prediction. Normally a model writes one token at a time, which is slow. An MTP head lets it guess the next few tokens in a burst, then the model checks the guesses itself and keeps only the right ones. Wrong guesses cost nothing. Right guesses are free speed.

So Fable performed the graft: 20 MTP tensors from Qwen, one extra transformer layer plus the nextn projection and norms, appended onto the quantized Ornith weights as block 41. Then it patched the metadata, bumped the block count from 40 to 41 and set nextn_predict_layers to 1. Brain surgery, on a GGUF file, on a Saturday.

The results, measured with llama.cpp build b9859 at a draft depth of two:

• Code generation: 77.9 tokens per second, 39 percent faster
• Prose generation: 71.2 tokens per second, 27 percent faster

And the part that matters: absolutely no drop in quality. This is output exact speculative decoding. The graft only lets Ornith guess what comes next. Ornith still verifies every guess itself and throws out the wrong ones, so the output is identical to the base model, token for token. Pure speed, zero tradeoff.

The whole thing fits real hardware. The Q4_K_M file is 21.7 GB and runs fully on GPU on any 24 GB card, a P40, a 3090 or a 4090, with a practical 128k context using a q8_0 KV cache. To our knowledge it is the only MTP build of this model that fits a standard consumer GPU.

We published the model on Hugging Face: Ornith-1.0-35B-MTP-Q4_K_M-GGUF. Licensing is clean, MIT from the Ornith base with the grafted tensors keeping their Apache 2.0. If you want to try it, this is the whole incantation:

llama-server -m ornith-1.0-35b-MTP-graft-Q4_K_M.gguf \
  -ngl 99 -c 131072 -ub 1024 \
  --cache-type-k q8_0 --cache-type-v q8_0 \
  -fa on --spec-type draft-mtp --spec-draft-n-max 2 \
  --jinja

It also loads in LM Studio, vLLM, Ollama and Docker. If you graft a head onto something of your own, tell us how it goes.