---

license: apache-2.0

base_model: badtheorylabs/BTL-4

library_name: llama.cpp

pipeline_tag: text-generation

tags:

- gguf

- btl-4

- mixture-of-experts

- speculative-decoding

- dspark

- llama.cpp

- agentic

- code

- blackwell

- long-context

---

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

| File | Purpose | Size |

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

| BTL-4-IQ2_XXS.gguf | verifier/target model | 9.97 GB |

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

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

| 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 | 260.51, 267.85, 271.10 | 266.49 tok/s | 271.10 | 12,586 MiB | 1.000x |

| DSpark, max draft 3 | 275.16, 307.40, 312.49 | 298.35 tok/s | 312.49 | 17,264 MiB | 1.120x |

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

| 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 | IQ2_XXS experts / Q4_K_M mixture, 9.97 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 and quantization caveat

The upstream full-precision

badtheorylabs/BTL-4 model card

reports the following official-harness results:

| Benchmark | Upstream BTL-4 BF16 | Attribution |

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

| LiveCodeBench v6 | 66.1% pass@1 | Bad Theory Labs; 442 problems, 2024-08 through 2025-05 |

| SWE-bench Verified | 78.4% | Bad Theory Labs; official harness |

| BFCL v4 AST | 73.5% | Bad Theory Labs; 1,240 cases |

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.