moontato/occamy-1.0-mtp-gguf

Occamy 1.0 MTP — GGUF Draft Heads

Community GGUF conversions of the experimental Occamy 1.0 MTP draft head from Accio Lab, intended for speculative decoding with Occamy 1.0 in llama.cpp.

These files are not standalone language models. They contain the MTP draft head and must be used together with a compatible Occamy 1.0 target model.

This repository is an unofficial community conversion and is not affiliated with or endorsed by Accio Lab, Qwen, or the llama.cpp project.

Upstream models

• Occamy 1.0 MTP: https://huggingface.co/Accio-Lab/occamy-1.0-MTP
• Occamy 1.0: https://huggingface.co/Accio-Lab/occamy-1.0
• Qwen3.6-35B-A3B: https://huggingface.co/Qwen/Qwen3.6-35B-A3B
• llama.cpp: https://github.com/ggml-org/llama.cpp

The upstream Occamy 1.0 MTP release is an experimental MTP head adapted on Occamy data. Accio Lab documents it as a separate draft head rather than a standalone model and notes that its validated runtime configuration is limited; these GGUF conversions are therefore community-produced and should be treated as experimental.

Files

This repository contains the MTP drafter in several GGUF precisions:

File	Precision	Suggested use
occamy-1.0-mtp-bf16.gguf	BF16	Reference / highest-fidelity drafter
occamy-1.0-mtp-Q8_0.gguf	Q8_0	Lower memory use with high fidelity
occamy-1.0-mtp-Q4_K_M.gguf	Q4_K_M	Smaller experimental drafter

The exact filenames in the repository may differ slightly; select the corresponding precision when downloading.

Conversion

The GGUFs were produced from the official Accio Lab Occamy MTP head by first assembling it with the Occamy 1.0 Hugging Face checkpoint, then exporting only the MTP head with a recent llama.cpp converter.

Representative workflow:

python assemble_head.py \
  --base /path/to/occamy-1.0 \
  --head ./mtp-trained.safetensors \
  --out /path/to/occamy-with-mtp

python convert_hf_to_gguf.py \
  /path/to/occamy-with-mtp \
  --mtp \
  --outtype bf16 \
  --outfile occamy-1.0-MTP-BF16.gguf

The lower-precision files were then quantized from the GGUF conversion using llama.cpp tooling.

llama.cpp usage

Use the MTP GGUF as the draft model and your normal Occamy 1.0 GGUF as the target model.

Example:

llama-server \
  -m /path/to/occamy-1.0-target.gguf \
  -md /path/to/occamy-1.0-MTP-BF16.gguf \
  --spec-type draft-mtp \
  --spec-draft-n-max 2

Add the rest of your normal target-model options as needed, such as GPU offload, context size, KV-cache precision, multimodal projector, and sampling settings.

A recent llama.cpp build is recommended because MTP support and GGUF conversion support continue to evolve.

Sampling

For the Occamy target model, the upstream recommended sampling values are:

temperature        = 1.0
top_p              = 0.95
top_k              = 20
presence_penalty   = 1.5

Sampling parameters can affect speculative acceptance rates. Compare drafter precisions using the same target model, prompt, sampler settings, context, and llama.cpp build.

Local acceptance-rate observations

In local llama.cpp testing with the BF16 MTP drafter, observed draft acceptance varied substantially with workload:

• approximately 50–75% acceptance across tested prompts
• coding and general-purpose tasks produced different acceptance behavior
• results should be considered anecdotal rather than a benchmark

Acceptance can vary with:

• prompt and workload
• target-model quantization
• MTP drafter quantization
• sampling parameters
• context length
• --spec-draft-n-max
• llama.cpp version / commit
• backend and hardware

The BF16 file is included as a useful reference when measuring any loss in acceptance caused by Q8_0 or Q4_K_M quantization.

Reproducibility

If publishing your own results, it is helpful to include:

llama.cpp commit/build:
Target GGUF:
Draft GGUF:
Context length:
KV cache type:
Temperature:
Top-p:
Top-k:
Presence penalty:
--spec-draft-n-max:
Prompt/workload:
Drafted tokens:
Accepted tokens:
Acceptance rate:
Baseline tok/s:
MTP tok/s:

Acceptance percentage alone does not necessarily imply a proportional throughput improvement; measure wall-clock generation speed as well.

Important caveats

• These GGUF files are MTP draft heads, not complete models.
• They are intended to be paired with Occamy 1.0, not arbitrary Qwen-family targets.
• The upstream MTP head is experimental.
• Accio Lab's published validation does not constitute validation of these community GGUF conversions.
• Quantizing a speculative drafter can reduce acceptance. BF16 is the best reference point for determining whether a lower-precision draft is worthwhile on your hardware.
• MTP support in llama.cpp is actively developed, so behavior may change between builds.

Provenance

Model lineage:

Qwen/Qwen3.6-35B-A3B
        ↓
Accio-Lab/occamy-1.0
        ↓
Accio-Lab/occamy-1.0-MTP
        ↓
Community GGUF conversion / quantization

No claim is made that these files are official Accio Lab GGUF releases.

License

The upstream Qwen3.6-35B-A3B, Occamy 1.0, and Occamy 1.0 MTP repositories are published under Apache License 2.0. This repository includes a copy of the Apache License 2.0 in LICENSE.

Please retain upstream attribution and comply with the terms and notices of the original works. If the exact upstream revisions used for your conversion contain a NOTICE file or other required attribution notices, those should also be preserved in redistributed derivatives.

Apache License 2.0: https://www.apache.org/licenses/LICENSE-2.0

Credits

• Accio Lab — Occamy 1.0 and the Occamy 1.0 MTP head
• Qwen Team — Qwen3.6-35B-A3B base model
• llama.cpp contributors — GGUF conversion, quantization, and speculative decoding support