GGUF conversions of AEON-7/Ornith-1.0-35B-AEON-Ultimate-Uncensored with MTP (Multi-Token Prediction) tensors grafted in for speculative decoding support.

The base model is the AEON abliterated variant of deepreinforce-ai/Ornith-1.0-35B. The MTP tensors are from unsloth/Qwen3.6-35B-A3B-MTP-GGUF.

Available Quantizations

File	Quant	Size	Source Weights	MTP Source
`Ornith-1.0-35B-AEON-Ultimate-Uncensored-MTP-BF16.gguf`	BF16	71.07 GB (66.19 GiB)	AEON-7/...-BF16	unsloth/Qwen3.6-35B-A3B-MTP-GGUF
`Ornith-1.0-35B-AEON-Ultimate-Uncensored-MTP-NVFP4.gguf`	NVFP4	23.40 GB (21.80 GiB)	AEON-7/...-NVFP4	unsloth/Qwen3.6-35B-A3B-MTP-GGUF via s-batman
`Ornith-1.0-35B-AEON-Ultimate-Uncensored-MTP-Q8_0.gguf`	Q8_0	37.80 GB (35.21 GiB)	Quantized from the BF16 GGUF	unsloth/Qwen3.6-35B-A3B-MTP-GGUF

Checksums

text

a13df4cce8a32b2065d8aea51dcc80d7056fea6c3277266d9b040923a2641840  Ornith-1.0-35B-AEON-Ultimate-Uncensored-MTP-BF16.gguf
d78f62f6c112de9721390ce8f75b22cf753b3766a640257cc13ca85f16030292  Ornith-1.0-35B-AEON-Ultimate-Uncensored-MTP-NVFP4.gguf
16696fb2e19b5b3faa316b198524be3dff3652555c67c3f3ea11e811147b219a  Ornith-1.0-35B-AEON-Ultimate-Uncensored-MTP-Q8_0.gguf

Provenance

Base model

text

deepreinforce-ai/Ornith-1.0-35B
  -> AEON-7/Ornith-1.0-35B-AEON-Ultimate-Uncensored-BF16
      -> AEON-7/Ornith-1.0-35B-AEON-Ultimate-Uncensored-NVFP4

MTP tensors

All MTP prediction heads originate from unsloth/Qwen3.6-35B-A3B-MTP-GGUF. These are compatible at the tensor-shape level because Ornith-1.0-35B uses the Qwen3.5 MoE architecture and tokenizer family.

BF16: blk.40.* MTP tensors grafted directly from Unsloth's BF16 split GGUF.
NVFP4: blk.40.* MTP tensors grafted via s-batman/Ornith-1.0-35B-NVFP4-MTP-GGUF. Byte-level verification confirms this block is identical to Unsloth's Qwen3.6-35B-A3B-MXFP4_MOE.gguf MTP block: 20 tensors, 512,079,872 tensor payload bytes, combined tensor-name-plus-payload SHA-256 8b8ba06cf776d2cdbf4d4db6714cf69b8a455105fc848bc02c4e5acb62f585f1.
Q8_0: blk.40.* MTP tensors grafted directly from Unsloth's Q8_0 GGUF.

Credit for the Qwen3.6 MTP tensors goes to Unsloth and the original Qwen release. s-batman is acknowledged as the intermediary who performed the NVFP4 graft used here as the practical donor for the NVFP4 file.

Usage

MTP requires a llama.cpp build with draft-mtp speculative decoding support.

llama-cli

bash

llama-cli \
  -m Ornith-1.0-35B-AEON-Ultimate-Uncensored-MTP-Q8_0.gguf \
  --spec-type draft-mtp \
  --spec-draft-n-max 3 \
  -p "Explain gradient descent in 3 sentences."

llama-server

bash

llama-server \
  -m Ornith-1.0-35B-AEON-Ultimate-Uncensored-MTP-NVFP4.gguf \
  --host 0.0.0.0 --port 8080 \
  -ngl all \
  -c 65536 \
  --spec-type draft-mtp \
  --spec-draft-n-max 3

Some frontends expose this as "MTP" or "speculative decoding" rather than the raw llama.cpp --spec-type draft-mtp flag.

LM Studio

Download the desired quant file. The model should appear as:

text

mrexodia/Ornith-1.0-35B-AEON-Ultimate-Uncensored-MTP-GGUF:BF16
mrexodia/Ornith-1.0-35B-AEON-Ultimate-Uncensored-MTP-GGUF:NVFP4
mrexodia/Ornith-1.0-35B-AEON-Ultimate-Uncensored-MTP-GGUF:Q8_0

Enable MTP/speculative decoding in advanced settings if your frontend supports it.

Notes on correctness

MTP is used as a speculative draft. The target model verifies proposed tokens, so a poorly matched MTP head should reduce acceptance rate or speedup rather than change the final verified output distribution. The graft is still experimental and should be benchmarked for your workload.

NVFP4 is intended for hardware and software stacks with NVFP4 support. On unsupported hardware, use the BF16 or Q8_0 files.

Exact production steps

All commands below were run from a llama.cpp checkout with a CUDA build available. Local cache paths are omitted for readability and shown as HuggingFace repo names.

1. Convert AEON BF16 safetensors to body-only BF16 GGUF

The source config advertises MTP, but the AEON BF16 safetensors snapshot does not contain MTP tensors. The body conversion was therefore done with --no-mtp.

bash

python convert_hf_to_gguf.py \
  AEON-7/Ornith-1.0-35B-AEON-Ultimate-Uncensored-BF16 \
  --outtype bf16 \
  --no-mtp \
  --outfile Ornith-1.0-35B-AEON-Ultimate-Uncensored-BF16.gguf

2. Graft BF16 MTP tensors

Copied all 20 donor tensors with prefix blk.40. from:

unsloth/Qwen3.6-35B-A3B-MTP-GGUF/BF16/Qwen3.6-35B-A3B-BF16-00002-of-00002.gguf

into the BF16 body GGUF, then updated:

text

qwen35moe.block_count = 41
qwen35moe.nextn_predict_layers = 1

The final published filename is:

Ornith-1.0-35B-AEON-Ultimate-Uncensored-MTP-BF16.gguf

3. Convert AEON NVFP4 safetensors to body-only NVFP4 GGUF

bash

python convert_hf_to_gguf.py \
  AEON-7/Ornith-1.0-35B-AEON-Ultimate-Uncensored-NVFP4 \
  --outtype bf16 \
  --no-mtp \
  --outfile Ornith-1.0-35B-AEON-Ultimate-Uncensored-NVFP4.gguf

The converter detected and preserved the source NVFP4 quantization. --outtype bf16 only affects non-NVFP4 tensors that remain floating point.

4. Graft NVFP4/MXFP4_MOE MTP tensors

Copied all 20 donor tensors with prefix blk.40. from:

s-batman/Ornith-1.0-35B-NVFP4-MTP-GGUF/ornith-1.0-35b-NVFP4_MOE-MTP.gguf

This block was verified byte-for-byte identical to the MTP block in:

unsloth/Qwen3.6-35B-A3B-MTP-GGUF/Qwen3.6-35B-A3B-MXFP4_MOE.gguf

Then updated:

text

qwen35moe.block_count = 41
qwen35moe.nextn_predict_layers = 1

The final published filename is:

Ornith-1.0-35B-AEON-Ultimate-Uncensored-MTP-NVFP4.gguf

5. Quantize BF16 body to Q8_0

The Q8_0 trunk was quantized from the body-only BF16 GGUF:

bash

llama-quantize \
  Ornith-1.0-35B-AEON-Ultimate-Uncensored-BF16.gguf \
  Ornith-1.0-35B-AEON-Ultimate-Uncensored-Q8_0-body.gguf \
  q8_0

6. Graft Q8_0 MTP tensors

Copied all 20 donor tensors with prefix blk.40. from:

unsloth/Qwen3.6-35B-A3B-MTP-GGUF/Qwen3.6-35B-A3B-Q8_0.gguf

Then updated:

text

qwen35moe.block_count = 41
qwen35moe.nextn_predict_layers = 1

The final published filename is:

Ornith-1.0-35B-AEON-Ultimate-Uncensored-MTP-Q8_0.gguf

7. Metadata updates

Each final GGUF was rewritten with file-specific metadata:

general.name
general.author = mrexodia
general.quantized_by = mrexodia
general.license = mit
general.license.name = MIT License
general.license.link = https://huggingface.co/deepreinforce-ai/Ornith-1.0-35B/blob/main/LICENSE
general.source.huggingface.repository
general.description
general.base_model.count = 2
general.base_model.0.* for the AEON source
general.base_model.1.* for the Unsloth MTP donor
general.tags

The legacy custom key general.base_model was removed in favor of the interoperable general.base_model.{id}.name mapping used by HuggingFace GGUF metadata.

8. Verification

For each final GGUF:

qwen35moe.block_count = 41
qwen35moe.nextn_predict_layers = 1
20 tensors with prefix blk.40. are present
4 tensors under blk.40.nextn.* are present
llama.cpp loaded the model with --spec-type draft-mtp
Smoke test prompt What is 2+2? Answer with just the number. produced 4

License

MIT, inherited from the base model.

AninoDevel/Ornith-1.0-35B-AEON-Ultimate-Uncensored-MTP-GGUF

Ornith-1.0-35B-AEON-Ultimate-Uncensored MTP GGUF

Available Quantizations

Checksums

Provenance

Base model

MTP tensors

Usage

llama-cli

llama-server

LM Studio

Notes on correctness

Exact production steps

1. Convert AEON BF16 safetensors to body-only BF16 GGUF

2. Graft BF16 MTP tensors

3. Convert AEON NVFP4 safetensors to body-only NVFP4 GGUF

4. Graft NVFP4/MXFP4_MOE MTP tensors

5. Quantize BF16 body to Q8_0

6. Graft Q8_0 MTP tensors

7. Metadata updates

8. Verification

License

AninoDevel/
Ornith-1.0-35B-AEON-Ultimate-Uncensored-MTP-GGUF