Same NVFP4 expert layout as our Qwen3.6-35B-A3B-NVFP4-GGUF repo, plus the MTP draft head extracted from the source for use with --spec-type draft-mtp.

Honest note up front: on the RTX 5090 with current llama.cpp, MTP does not yet win on this MoE for single-stream generation — the base MoE is already so fast (~250 tok/s with NVFP4-Q4_K_M) that draft + verify overhead exceeds the savings. The tuned config below gets it to break-even (-2%); the default loses ~27%. We're publishing the files anyway because (a) the MTP path is rapidly being optimized upstream and will likely flip, and (b) MTP weights are useful for downstream training/research. See the Performance section for numbers.

About LibertAI

LibertAI is a decentralized AI platform — private inference, an OpenAI-compatible API, and a chat UI, all running on community GPUs over Aleph Cloud instead of a single company's servers.

If you want to put this model (or any other) to work as an autonomous agent without running your own infrastructure, check out LiberClaw — Hermes-style agents hosted on Aleph Cloud with LibertAI inference.

Files

File	Size	Experts	Other tensors	When to pick
`Qwen3.6-35B-A3B-NVFP4-Q4_K_M-mtp.gguf`	19 GB	NVFP4	Q4_K_M	Recommended trunk
`Qwen3.6-35B-A3B-NVFP4-Q8_0-mtp.gguf`	20 GB	NVFP4	Q8_0	Higher precision attention/embeddings
`Qwen3.6-35B-A3B-NVFP4-BF16-mtp.gguf`	22 GB	NVFP4	BF16	Max source-fidelity for non-expert tensors
`mtp-Qwen3.6-35B-A3B-NVFP4.gguf`	3.5 GB	BF16 (MoE)	BF16/F32	MTP draft head — required for `--spec-type draft-mtp`. Itself a small MoE block (256 experts, 1 layer) plus shared experts and the `nextn` heads
`mmproj-Qwen3.6-35B-A3B-F16.gguf`	889 MB	—	F16 vision tower	Required for image/video input

The trunks are built with convert_hf_to_gguf.py --no-mtp. The MTP weights are split into the separate mtp-*.gguf file so they can be loaded as a draft via --model-draft.

Performance

Measured on an NVIDIA RTX 5090 (32 GB, Blackwell, sm_120, CUDA 13.0), llama.cpp build dbe9c0c8c.

Single-stream serving, 28-token prompt → 512-token completion (n=3, prod settings: -c 262144 -fa on -ctk q4_0 -ctv q4_0 -ub 256 -b 1024 --parallel 1):

Config	TG (tok/s)	Draft accept rate	Δ vs no-MTP
NVFP4-Q4_K_M, no MTP	252.9	—	baseline (fastest)
NVFP4-Q4_K_M + MTP (n-max=2, p-min=0.3)	248.6	63.5%	−2%
NVFP4-Q4_K_M + MTP (n-max=4, p-min=0.5, default)	184.8	65.6%	−27%

The MoE base is so fast that MTP currently doesn't pay for itself at single-stream. With the tuned aggressive config (n-max=2 p-min=0.3), we get within 2% of baseline. We expect this to flip once the upstream llama.cpp MoE+MTP code path is optimized — the same was true for plain NVFP4 on MoE before recent improvements landed.

Until then: if you don't have a strong reason to use MTP on this MoE, prefer the non-MTP repo.

Usage

Server with MTP (experimental; tuned config)

bash

llama-server \
  -m Qwen3.6-35B-A3B-NVFP4-Q4_K_M-mtp.gguf \
  --model-draft mtp-Qwen3.6-35B-A3B-NVFP4.gguf \
  --spec-type draft-mtp \
  --spec-draft-n-max 2 \
  --spec-draft-p-min 0.3 \
  -ngl 999 -ngld 999 \
  -fa on -c 32768 \
  --host 0.0.0.0 --port 8080

Server without MTP (currently fastest single-stream)

bash

llama-server \
  -m Qwen3.6-35B-A3B-NVFP4-Q4_K_M-mtp.gguf \
  -ngl 999 \
  -fa on -c 32768 \
  --host 0.0.0.0 --port 8080

Multimodal

Add --mmproj mmproj-Qwen3.6-35B-A3B-F16.gguf to either of the above for image/video input.

Recommended sampler

Qwen3.6 is a thinking model. For non-thinking usage, set chat_template_kwargs.enable_thinking=false in the API.

What is MTP?

Multi-Token Prediction adds a small extra block (here, 1 MoE layer at index 40) that produces a candidate next-token distribution from the same hidden state as the trunk. llama.cpp uses this as a draft for speculative decoding: generate up to N candidates, verify them in a single trunk forward pass, keep the accepted ones.

The MTP path landed in llama.cpp via PR #22673; NVFP4 scale-tensor support for the MTP block in MoE models landed in #23563.

Architecture notes

Qwen3.6-35B-A3B is an MoE model with 256 routed experts (8 active per token) plus a shared expert, hybrid attention + SSM (full attention every 4th of 40 layers). The NVFP4 source from mmangkad quantizes the routed expert FFN (ffn_{down,gate,up}_exps, 120 tensors = 40 layers × 3) — attention, SSM, and shared experts stay at higher precision. The MTP block at layer 40 is itself a small MoE layer with the same 256-expert structure (kept BF16 since the source didn't quantize it).

Sources & credits

Base model: Qwen/Qwen3.6-35B-A3B by Alibaba Qwen team — Apache 2.0
NVFP4 calibration source: mmangkad/Qwen3.6-35B-A3B-NVFP4 (NVIDIA ModelOpt v0.43)
mmproj source: official BF16 weights from Qwen/Qwen3.6-35B-A3B
Tooling: llama.cpp convert_hf_to_gguf.py --no-mtp / --mtp and llama-quantize

License

Apache 2.0, inherited from the upstream model.

LibertAIDAI/Qwen3.6-35B-A3B-NVFP4-MTP-GGUF

Qwen3.6-35B-A3B NVFP4 GGUF — MTP variant