jcbtc/qwen3.6-35b-a3b-crown-halo-mtp-dynamic

Qwen3.6 35B A3B Crown Halo Dynamic MTP v7 GGUF

[image]

Crown Dynamic MTP v7 is a custom mixed-precision GGUF quantization of unsloth/Qwen3.6-35B-A3B-MTP-GGUF, tuned for local AMD Strix Halo serving with llama.cpp native MTP speculative decoding.

This is not a generic cloud inference recipe. It is a Strix Halo owner profile: single-user, high-context, Vulkan, native MTP, split mode none, flash attention, vision projector support, and polling settings chosen for this machine class.

File

mmproj-F16.mmproj is the Qwen3.6 35B A3B GGUF-format vision projector. It is stored with a .mmproj repo extension so Hugging Face's GGUF parser keeps the main card focused on the 35B language model rather than the smaller projector.

Technical Metadata

Hugging Face may round the parsed GGUF tensor count to 36B in its automatic badge. This release is the Qwen3.6 35B-A3B MoE family: about 35B-class total parameters with roughly 3B active parameters per token.

What This Is

Crown v7 is not a renamed upstream quant. It uses a custom mixed-precision tensor recipe:

• —MXFP4 on blocks 0-27 routed expert gate/up tensors
• —Q4_K_M fallback on blocks 28-35 routed expert gate/up tensors
• —Q5_K on blocks 36-39 routed expert gate/up tensors
• —Q5_K / Q6_K for routed down experts
• —Q8_0 for attention, shared experts, token embeddings, output, and selected MTP tensors

The full recipe is included in recipes/halo-mtp-dyn-v7.md and recipes/halo-mtp-dyn-v7.tensor-types.txt.

Strix Halo Serving Profile

Reference profile:

• —AMD Ryzen AI MAX / Radeon 8060S Strix Halo class machine
• —Vulkan llama.cpp build with Qwen MTP support
• —128 GB unified memory
• —single-slot serving because MTP is the constraint
• —131072 context
• —split mode none
• —flash attention enabled
• —native MTP speculative decoding
• —Qwen3.6 35B A3B vision projector enabled

Primary Strix/MTP/vision showcase command:

llama-server \
  -m Qwen3.6-35B-A3B-HaloStrix-Dyn-MTP-v7.gguf \
  --alias crown-dynamic-mtp \
  --host 127.0.0.1 \
  --port 18181 \
  --jinja \
  -c 131072 \
  --reasoning off \
  --reasoning-format none \
  --reasoning-budget -1 \
  --no-context-shift \
  -sm none \
  -ngl 999 \
  -fa on \
  -b 2048 \
  -ub 512 \
  -t 16 \
  -ctk f16 \
  -ctv f16 \
  --spec-draft-type-k f16 \
  --spec-draft-type-v f16 \
  --parallel 1 \
  --metrics \
  --mmproj mmproj-F16.mmproj \
  --spec-type mtp \
  --spec-draft-n-max 4 \
  --poll 100 \
  --poll-batch 1 \
  --spec-draft-poll 1 \
  --spec-draft-poll-batch 1 \
  --temp 0.6 \
  --min-p 0.0 \
  --top-p 0.95 \
  --top-k 20 \
  --repeat-penalty 1.0

The exact public Strix profile is included as model-profiles/qwen3.6-35b-a3b-crown-halo-mtp-dynamic.env, and the serving script is included as scripts/serve_halo_mtp_dyn_v7.sh.

For text-only use, you may omit --mmproj. For image input, keep mmproj-F16.mmproj next to the main GGUF and pass it with --mmproj.

KV Cache Choice

Use f16 target KV and f16 draft KV with b2048/u512.

Older local comparison rows included q8_0/q8_0 KV, but that is not the public Strix Halo recommendation for this release. The benchmark story and serving recipe are centered on the f16/f16 MTP profile.

Where It Shines On Strix Halo

The point of this release is native MTP behavior under compatible llama.cpp serving. It shines when the MTP head gets accepted at a high rate, especially structured or repetitive long-decode workloads.

On structured 4k-token prompt / 256-token generation tests:

Best observed structured-decode slice: MTP depth 4, roughly 105-106 tok/s with 97-99% draft-token acceptance.

A separate context sweep found a strong mid-context decode row:

Vision Smoke Test

Vision was validated locally on June 15, 2026 with the normal Crown Dynamic profile, the mmproj-F16 projector, and native MTP enabled. Test image:

/srv/desktop-data/cirudata/ciruoutfit.png

Prompt:

Look at this image. In one short sentence, name the dominant hair color and the object covering the character's ear.

Response:

The dominant hair color is green, and the object covering the character's ear is a cybernetic headset.

The OpenAI-compatible llama.cpp response included MTP draft timing fields: draft_n=24 and draft_n_accepted=19.

Strix Owner Notes

• —--parallel 1 is intentional. MTP is the constraint, and this profile is for single-user local serving.
• —-c 131072 is intentional. This release is focused on high-context Strix Halo use.
• —Use --spec-type mtp; plain GGUF loading will not show the behavior this model was built for.
• —Use the polling settings from the profile: --poll 100, --poll-batch 1, --spec-draft-poll 1, and --spec-draft-poll-batch 1.
• —b2048/u512 is the recommended Strix MTP default from the acceptance matrix; larger b8192/u512 did not improve acceptance and added latency.
• —f16/f16 KV is the recommended MTP setting for Strix Halo.
• —Use --mmproj mmproj-F16.mmproj for image input. Omit it only for text-only serving.

Important Benchmark Caveat

This is not a universal "always faster" GGUF.

In ordinary non-speculative llama-bench, Crown v7 looks similar to other strong Qwen3.6 35B A3B quants. The point of this release is the MTP behavior under compatible Strix Halo llama.cpp serving, not raw no-speculative token generation.

Local context-sweep testing showed:

• —MTP helped most on structured/repetitive decode workloads with high draft acceptance.
• —MTP depth 4 was the strongest tested setting for structured decode.
• —Very long active contexts, especially around 64k, did not show an end-to-end MTP win on this Vulkan host.
• —If you run this as a normal GGUF without MTP support, expect a strong Q4-size mixed quant, not the accelerated MTP profile.

Provenance

Base/source repo: unsloth/Qwen3.6-35B-A3B-MTP-GGUF

Quantization recipe used BF16 split GGUF shards and imatrix_unsloth.gguf_file from the upstream Unsloth release.

Recipe checksums:

• —recipes/halo-mtp-dyn-v7.md: 9b56f075cbd5cc225c28ae7022c6c0e04beafeec6fa275f50aef4c0d1d2f46cc
• —recipes/halo-mtp-dyn-v7.tensor-types.txt: d827103057fb485db6c4e7116bdb3658a50b48a79c915431f18188bb647d2375