Ornith-1.0-9B MTP — GGUF (llama.cpp speculative decoding)

GGUF builds of deepreinforce-ai/Ornith-1.0-9B with the KL-distilled MTP draft head from protoLabsAI/Ornith-1.0-9B-MTP baked into the trunk — llama.cpp does lossless multi-token self-speculative decoding out of the box, no separate draft model to wire up. Every file here carries the nextn head, so --spec-type draft-mtp just works.

Two things to know before you pick a file:

• Ampere and older: use Q4_K_M. It is smaller and faster than everything else here.
• Blackwell (RTX 50xx / PRO 6000): use NVFP4. MTP compounds with NVFP4's tensor-core GEMMs where it only partially helps K-quants — so NVFP4 is the fastest rung on this hardware by ~28%, despite being a hair larger than Q4_K_M. The measured mechanism is below.

Want the base with no MTP head? deepreinforce-ai/Ornith-1.0-9B-GGUF.

Files

"Bundled" = trunk + nextn head in one file. The IQ rungs are i-quants (importance-matrix calibrated) with the MTP head pinned to Q8_0 so acceptance holds on the low-bit trunk (measured ~0.81–0.84 on IQ2_M–IQ4_XS, on par with the k-quants). Serve them exactly like the k-quants.

The standalone head is not a model — loading mtp-head/… directly will crash. It exists only to pair with a base Ornith-9B GGUF via --model-draft.

Requires llama.cpp ≥ b9616 (Qwen3.5 qwen35 arch + --spec-type draft-mtp). The NVFP4 rung additionally needs a build with NVFP4 support (GGML_TYPE_NVFP4, type 40) — both landed in llama.cpp spring 2026. LM Studio (recent) and Ollama (≥ ~0.31) inherit MTP through llama.cpp; older Ollama fails with layer 32 missing attn_qkv → update and re-pull.

Run

Bundled (recommended) — the head travels in the file:

llama-server --model Ornith-1.0-9B-MTP-Q4_K_M.gguf \
  --n-gpu-layers 99 --ctx-size 8192 --flash-attn on --jinja \
  --spec-type draft-mtp --spec-draft-n-max 3

One-command pull:

llama-server -hf protoLabsAI/Ornith-1.0-9B-MTP-GGUF:NVFP4  --spec-type draft-mtp -ngl 99  # Blackwell
llama-server -hf protoLabsAI/Ornith-1.0-9B-MTP-GGUF:Q4_K_M --spec-type draft-mtp -ngl 99  # everything else

Standalone draft — pair the small head with any base Ornith-9B GGUF:

llama-server --model ornith-1.0-9b-Q4_K_M.gguf \
  --model-draft mtp-head/mtp-Ornith-1.0-9B-head-Q8_0.gguf \
  --spec-type draft-mtp --spec-draft-n-max 3 \
  --n-gpu-layers 99 --ctx-size 8192 --flash-attn on --jinja

--spec-draft-n-max is the draft depth: 2 maximizes acceptance, 3 maximizes throughput, 4 starts to regress. Tune per workload.

The NVFP4 × MTP finding (why Blackwell is different)

NVFP4 weights sit on the tensor-core FP4 GEMM path; K-quants dequantize to a compute path each step. MTP's cost is a per-step parallel verify of the drafted tokens — and that verify is nearly free on NVFP4's GEMMs but costs ~+28% on the K-quant dequant path. So MTP's speedup is effectively multiplicative with NVFP4 and only partial with K-quants.

Measured, acceptance-controlled (-n 200 greedy, 6 diverse prompts, ranges shown; Ampere single-run indicative):

file        size    Ampere A6000        Blackwell (sm120)
                    no-MTP   +MTP       no-MTP   +MTP
---------   ------  ------   ------     ------   ----------------
Q4_K_M      5.8 GB  104.6    153.4      205.1    239  (216–252)
NVFP4       6.6 GB   70.7     84.8      201.5    306  (287–330)

On Blackwell the worst NVFP4 prompt (287) beats the best Q4_K_M prompt (252); draft acceptance is near-equal on both files at the same prompt/box, so the gap is verify cost, not acceptance. On Ampere the FP4 path has no tensor-core backing, so Q4_K_M wins outright — use it. Code prompts run hottest (330), creative prose lowest (287): MTP acceptance tracks predictability.

The FP8 × spec-decode compounding is documented upstream (TensorRT-LLM); the FP4-vs-K-quant verify-cost split is, as far as we've found, new data.

Benchmarks (RTX A6000, ctx 8192, flash-attn, greedy; 6-prompt code+general mix)

n-max sweep, Q8_0

Per-token acceptance at n-max 2 (0.766) matches the vLLM reference for this head (0.762).

Across quants, n-max 3

Acceptance is quant-stable (~0.65 at n-max 3 even with a Q4 trunk). Q4_K_M is fastest in absolute terms on Ampere; the relative MTP gain grows with precision (Q8's bandwidth-bound baseline has more to gain from the parallel verify).

Full rows behind every number: protoLabsAI/lab-benchmarks · charts at protolabs.studio/lab.

"Lossless" — read this

MTP speculative decoding is distribution-lossless: every drafted token is verified against the target, so the output distribution is unchanged. It is not bitwise-identical to plain decode at greedy/temp 0 — the batched verify computes target logits in a different floating-point reduction order than sequential decode, which can flip a greedy argmax and fork the text. Both outputs are equally valid; this is expected llama.cpp behavior, not a defect of these weights.

Troubleshooting: wrong number of tensors expected 442 got 427

(or got 426 on smaller quants — the gap is the 15 mtp.* head tensors.)

This happens if you convert the base deepreinforce-ai/Ornith-1.0-9B directly without grafting the head first. The base keeps mtp_num_hidden_layers: 1 in config.json but ships none of the mtp.* weights, so the converter declares a blk.32 MTP layer while leaving those 15 tensors empty → llama.cpp expects 442 and finds 427.

Fix: graft the head into the trunk before converting, then convert with no --mtp flag. (Only 4 of the 15 head tensors are named blk.32.nextn.*; the other 11 land as ordinary blk.32.* tensors, so grep nextn shows 4 but the head is complete.) Or skip grafting entirely and run the base GGUF with --model-draft mtp-head/mtp-Ornith-1.0-9B-head-Q8_0.gguf — functionally identical.

How these were built

# 1. graft the mtp.* head into the base trunk (15 tensors, 1 nextn layer)
python graft.py --donor protoLabsAI/Ornith-1.0-9B-MTP \
                --target deepreinforce-ai/Ornith-1.0-9B --out ./ornith-9b-mtp-kl
# 2. convert (remaps mtp.* -> blk.32.nextn.* automatically)
python convert_hf_to_gguf.py ./ornith-9b-mtp-kl --outfile out/...-BF16.gguf --outtype bf16
# 3. quantize (NVFP4 rung converted from the gate-verified vLLM NVFP4 quant, same scales)
llama-quantize out/...-BF16.gguf out/...-Q4_K_M.gguf Q4_K_M

The graft.py recipe and the KL-distillation details live in the head repo protoLabsAI/Ornith-1.0-9B-MTP. The NVFP4 rung is converted from protoLabsAI/Ornith-1.0-9B-NVFP4 — full quality/coherence receipts on that card.

Want a different size or format?

Open a Community discussion — requests usually ship within 48h. That's how most of the quants here got made.

Provenance & license

• Base: deepreinforce-ai/Ornith-1.0-9B (MIT) — a Qwen3.5-9B hybrid (linear + full attention) fine-tune.
• MTP head: protoLabsAI/Ornith-1.0-9B-MTP (MIT) — KL-distilled against Ornith's own hidden states.
• These GGUFs derive from both; MIT. Built by protoLabs.studio.