---

license: apache-2.0

base_model: HauhauCS/Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive

library_name: llama.cpp

pipeline_tag: text-generation

tags:

- gguf

- qwen3.6

- mixture-of-experts

- speculative-decoding

- dspark

- llama.cpp

- blackwell

- long-context

---

Qwen3.6 35B A3B Aggressive Q4_K_M + DSpark GGUF

A self-contained, tested pairing of the Q4_K_M target from

HauhauCS/Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive

and the Qwen3.6 DSpark GGUF draft converted by

williamliao/Qwen3.6-35B-A3B-DSPARK-GGUF.

The weights are unchanged from those upstream releases. This repository adds a

reproducible llama.cpp launch profile and measured RTX 5090 results at a

200,704-token configured context window.

> DSpark support is experimental. These results use

> llama.cpp PR #26275,

> commit f806441edb5006fdacb07df42445f337476dd169.

Files

| File | Purpose | Size |

|---|---|---:|

| Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive-Q4_K_M.gguf | verifier/target model | 21.17 GB |

| Qwen3.6-35B-A3B-DSPARK.gguf | DSpark draft model; not standalone | 1.04 GB |

| launch-dspark.sh | tested 200K launch profile | — |

| prime-agent-models.json | Prime Agent custom-provider template | — |

| benchmark-results.json | machine-readable measurements | — |

RTX 5090 generation benchmark

Measured locally on 2026-08-07. Each result is an end-to-end wall-clock OpenAI

Chat Completions request generating 512 tokens. Values are three independent

runs after loading the full 200,704-token context configuration.

| Configuration | Runs (tok/s) | Mean | Best | VRAM | Relative to baseline |

|---|---|---:|---:|---:|---:|

| No draft | 270.79, 276.17, 279.65 | 275.54 tok/s | 279.65 | 23,266 MiB | 1.000x |

| DSpark, max draft 3 | 306.16, 312.70, 318.27 | 312.38 tok/s | 318.27 | 27,944 MiB | 1.134x |

| DSpark, max draft 5 | 233.74, 256.70, 260.33 | 250.26 tok/s | 260.33 | 28,070 MiB | 0.908x |

| DSpark, max draft 7 | 207.63, 224.25, 226.90 | 219.59 tok/s | 226.90 | 28,196 MiB | 0.797x |

--spec-draft-n-max 3 is the clear choice on this target. In the 3 × 512-token

coding run, 337 of 520 drafted tokens were accepted (64.81%). A separate short

generation accepted 41 of 45 drafted tokens (91.11%). Acceptance depends on the

prompt and output distribution.

Test machine

| Component | Value |

|---|---|

| GPU | NVIDIA GeForce RTX 5090, 32,607 MiB |

| Driver / power limit | 595.84 / 575 W |

| OS / kernel | Ubuntu 26.04 LTS / Linux 7.0.0-28-generic x86_64 |

| Target quantization | Q4_K_M, 21.17 GB |

| Draft | BF16 GGUF, 1.04 GB |

| Configured context | 200,704 tokens |

| KV cache | Q8_0 K and V for both target and draft |

| Batch / microbatch | 2,048 / 512 |

| Parallel slots | 1 |

| Flash attention | enabled |

These are single-stream decode measurements, not prompt-processing throughput

or multi-user aggregate throughput. The first request can include warm-up

overhead, which is intentionally retained in the mean.

Quality benchmarks

DSpark is target-verifying speculative decoding: accepted draft tokens are

verified by the target, so it accelerates inference without substituting the

draft model's predictions for the target's decisions.

The upstream Aggressive model card does not publish LiveCodeBench v6 or

SWE-bench Verified scores, and those suites were not rerun for this packaging

work. No quality score is claimed here. Refer to the

HauhauCS model card

for the target's reported properties and safety behavior.

Run

Build the experimental branch with CUDA enabled:

git clone --branch dspark-speculators https://github.com/wjinxu/llama.cpp.git
cmake -S llama.cpp -B llama.cpp/build \
  -DGGML_CUDA=ON -DCMAKE_BUILD_TYPE=Release \
  -DCMAKE_CUDA_ARCHITECTURES=120a
cmake --build llama.cpp/build --config Release -j --target llama-server

Then, from this repository directory:

LLAMA_SERVER=/path/to/llama.cpp/build/bin/llama-server ./launch-dspark.sh

OpenAI-compatible API:

curl http://127.0.0.1:8080/v1/chat/completions \
  -H 'Content-Type: application/json' \
  -d '{
    "model": "qwen3.6-35b-a3b-aggressive-dspark",
    "messages": [{"role": "user", "content": "Write a Python LRU cache."}],
    "temperature": 0,
    "max_tokens": 512
  }'

For Prime Agent, copy prime-agent-models.json to

~/.prime/agent/models.json, change YOUR_SERVER_IP, then select provider

qwen36-dspark and model qwen3.6-35b-a3b-aggressive-dspark.

Attribution and limitations

• Target and quantization: HauhauCS
• Base model: Qwen/Qwen3.6-35B-A3B
• Draft conversion: williamliao/Qwen3.6-35B-A3B-DSPARK-GGUF
• Draft lineage: Koopah/Qwen3.6-35B-A3B-NVFP4-DSPARK
• DSpark research/runtime: DeepSpec and llama.cpp PR #26275

This target is explicitly uncensored. Review the upstream model card and use it

responsibly. This bundle is text-only because it does not include the optional

vision projector. See THIRD_PARTY_NOTICES.md and LICENSE.