credits: the MTP head is Qwen's own trained block, not nerkyor's — byte-verified, and name pahajokiconsulting as the donor it was copied from

README.md

@@ -25,7 +25,7 @@ tags:
 
 # Ornith-1.0-35B-A3B — MXFP4 + MTP (vision), for AMD RDNA4 / vLLM
 
-[`deepreinforce-ai/Ornith-1.0-35B`](https://huggingface.co/deepreinforce-ai/Ornith-1.0-35B) — a
+[`deepreinforce-ai/Ornith-1.0-35B`](https://huggingface.co/ornith-ai/Ornith-1.0-35B) — a
 **Qwen MoE model (A3B, ~3 B active parameters per token)** — quantized to **MXFP4** with a **grafted MTP
 (Multi-Token-Prediction) draft head**, packaged to run **out of the box on AMD Radeon RDNA4 (gfx1201)
 under vLLM** with lossless self-speculative decoding. Vision retained.
@@ -36,7 +36,10 @@ under vLLM** with lossless self-speculative decoding. Vision retained.
 - **Draft head — a cross-model graft.** The MoE MTP head (785 BF16 tensors, with per-expert weights) is
   transplanted from a **sibling Qwen-MoE checkpoint**,
   [`Capicua25x/Qwen3.6-35B-A3B-DSV4Pro-Thinking-Distill-MXFP4-Vision`](https://huggingface.co/Capicua25x/Qwen3.6-35B-A3B-DSV4Pro-Thinking-Distill-MXFP4-Vision)
-  (a DeepSeek-V4-Pro-Thinking distill of Qwen3.6-35B-A3B). Because the trunk and the donor share the
+  (a DeepSeek-V4-Pro-Thinking distill of Qwen3.6-35B-A3B). The head weights originate with
+  [`Qwen/Qwen3.6-35B-A3B`](https://huggingface.co/Qwen/Qwen3.6-35B-A3B) (Apache-2.0) and reach this build unchanged
+  via [`pahajokiconsulting/Qwen3.6-35B-A3B-MXFP4`](https://huggingface.co/pahajokiconsulting/Qwen3.6-35B-A3B-MXFP4) —
+  every checkpoint in that chain carries them byte-identical. Because the trunk and the donor share the
   **same Qwen MoE architecture** — identical hidden size and expert layout — the head transfers cleanly
   and drafts well on the Ornith trunk (acceptance below).
 - **Speculative decoding** — vLLM native `mtp` method, `num_speculative_tokens=3`. **Lossless**: the
@@ -116,10 +119,11 @@ Step 1 (the MXFP4 quantize) is scripted in [`quantize_mxfp4.py`](./quantize_mxfp
 script handles both a dense head and this MoE head.
 
 ## Credits
-- **DeepReinforce** — [`Ornith-1.0-35B`](https://huggingface.co/deepreinforce-ai/Ornith-1.0-35B), the base/trunk model (MIT).
-- **nerkyor** — [`Qwen3.6-35B-A3B-DSV4Pro-Thinking-Distill`](https://huggingface.co/nerkyor/Qwen3.6-35B-A3B-DSV4Pro-Thinking-Distill), which trained the MoE MTP head grafted here.
+- **DeepReinforce** — [`Ornith-1.0-35B`](https://huggingface.co/ornith-ai/Ornith-1.0-35B), the base/trunk model (MIT). *(The `deepreinforce-ai/` id now redirects to `ornith-ai/`.)*
+- **nerkyor** — [`Qwen3.6-35B-A3B-DSV4Pro-Thinking-Distill`](https://huggingface.co/nerkyor/Qwen3.6-35B-A3B-DSV4Pro-Thinking-Distill), the DeepSeek-V4-Pro-Thinking **trunk distill**. *(Corrected 2026-08-17: this card previously credited the distill with training the MTP head. It did not — the head is Qwen's, carried through unchanged. Verified by hash: `mtp.fc.weight` [2048, 4096] and the other `mtp.*` tensors are byte-identical between `Qwen/Qwen3.6-35B-A3B` and the distill.)*
 - **DeepSeek** — DeepSeek-V4-Pro-Thinking, the model that distill learned from.
-- **Qwen / Alibaba** — the Qwen MoE (A3B) architecture shared by both the trunk and the grafted head.
+- **pahajokiconsulting** — [`Qwen3.6-35B-A3B-MXFP4`](https://huggingface.co/pahajokiconsulting/Qwen3.6-35B-A3B-MXFP4) (Apache-2.0), the build the 785 `mtp.*` tensors were copied from verbatim: it quantizes the trunk but keeps `mtp.*` in BF16, with the experts relaid from stacked 3D into per-expert 2D.
+- **Qwen / Alibaba** — the Qwen MoE (A3B) architecture shared by trunk and head, **and the trained MTP head weights themselves** ([`Qwen/Qwen3.6-35B-A3B`](https://huggingface.co/Qwen/Qwen3.6-35B-A3B), Apache-2.0). The grafted head is Qwen's own trained block, not a third party's.
 - **vLLM** and **compressed-tensors** — serving stack and quantization format.
 - **Rob Smith** (`tcclaviger`) — the [RDNA4 vLLM base image](https://hub.docker.com/r/tcclaviger/vllm-rocm-mxfp4-nvfp4) that made gfx1201 serving possible. The [`capicua25x/vllm-rocm-rdna4`](https://hub.docker.com/r/capicua25x/vllm-rocm-rdna4) image this model runs on is a forward-port of his work — without it, none of this runs.