BTL-4-IQ2_XXS-DSPARK-GGUF

BTL-4 IQ2_XXS + Qwen3.6 DSpark GGUF

A self-contained, tested pairing of badtheorylabs/BTL-4-Compact and the Qwen3.6 DSpark GGUF draft converted by williamliao/Qwen3.6-35B-A3B-DSPARK-GGUF.

BTL-4 is Qwen3.6/Ornith-derived and proved compatible with this draft in the experimental llama.cpp DSpark verifier. The weights are unchanged from the upstream releases. This repository adds a reproducible launch profile and RTX 5090 measurements at a 200,704-token configured context window.

DSpark support is experimental. These results use llama.cpp PR #26275, commit f806441edb5006fdacb07df42445f337476dd169.

Files

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.

The DSpark run accepted 326 of 553 drafted tokens (58.95%) with mean accepted length 2.76. --spec-draft-n-max 3 is recommended; longer draft settings were already slower on the Qwen3.6 Aggressive target tested on the same machine.

Test machine

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 and quantization caveat

The upstream full-precision badtheorylabs/BTL-4 model card reports the following official-harness results:

Those values are upstream-reported BF16 scores, not measurements of this IQ2_XXS file. The Compact card reports 94.1% behavioral retention (111/118 on its quantization replay gate), but LiveCodeBench and SWE-bench Verified were not rerun for the compact quant. Do not present 66.1% or 78.4% as measured compact scores.

DSpark is target-verifying speculative decoding: accepted draft tokens are verified by BTL-4, so it accelerates this quantized target without replacing the target's token decisions.

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": "btl-4-compact-dspark",
    "messages": [{"role": "user", "content": "Refactor this function."}],
    "temperature": 0,
    "max_tokens": 512
  }'

Attribution and limitations

• Compact target: badtheorylabs/BTL-4-Compact
• Full-precision target and reported quality benchmarks: badtheorylabs/BTL-4
• Draft conversion: williamliao/Qwen3.6-35B-A3B-DSPARK-GGUF
• Draft lineage: Koopah/Qwen3.6-35B-A3B-NVFP4-DSPARK
• Runtime: llama.cpp PR #26275

The upstream compact build is text-only and has its MTP layer disabled. The external DSpark draft is separate from MTP. See THIRD_PARTY_NOTICES.md and LICENSE.