howtospark
models / btl3
?

badtheorylabs/BTL-3
badtheorylabs/BTL-3
bfloat16GQA 24:4262K ctxmultimodalapache-2.0
Sparks
Spec decode
Context33K / 262KKV
layer map
64 layerslinear attention×48full attention×16
in
out

Recipes (all recipes)

Worked deployments that run this model on DGX Spark — each shows how it fills each node's unified memory.

81*
1
256K ctx12.81 tok/s

BTL-3 (Qwen3.6-27B) — BF16

badtheorylabs' BTL-3 agentic coder (a LoRA on Qwen3.6-27B, ~27B dense) in BF16 on one DGX Spark, with the base model's native MTP head driving speculative decode — 12.8 tok/s single-stream, 2.9x the no-draft baseline, at the full 262K context.

via badtheorylabs/BTL-3 — model card + adapter_config.json
Optimize
Quantized from

Qwen/Qwen3.6-27B

Analyze the original repo for the full architecture picture.

Verdict

Couldn't read a parameter count — this may not be a standard safetensors transformer repo.

Make it fast

Once it fits: fewer bytes per token and fewer decode steps against 273 GB/s.

Add a speculative decoder

Decode is bandwidth-bound; a trained EAGLE-style draft multiplies tokens per weight-read.

EAGLEDraft-model speculative decodingMedusa heads
in the wild: vllm-Moet (kacper-daftcode)AngelSlim Eagle3 training (Tencent)

Tame the KV cache

At 262144 tokens the KV cache alone is ~69 GB at FP16; quantize or evict.

KV-cache quantizationToken eviction (H2O, StreamingLLM)MLA / GQA (architectural KV)
in the wild: vllm-Moet NVFP4 KV cache

Make it yours

Change what the model does — adapt, edit, or steer it — independent of size.

Fine-tune it

LoRA/QLoRA adapts behavior on one Spark without touching the base weights.

Supervised fine-tuning (SFT)Low-rank factorizationPreference tuning (RLHF / DPO)

Edit or steer it

One-shot weight math — no gradients, no corpus.

AbliterationControl vectors / steeringModel merging (mergekit)
in the wild: huihui-ai abliterated models