Open weights · Frozen release RL-0013

BTL–3

A 27B agentic coding and tool-use model trained to act, inspect, recover, and stop when no action is needed.

Post-trained from a pinned Qwen3.6-27B revision. Every result below belongs to the frozen RL-0013 adapter and carries its actual protocol and denominator.

Download weights ↗Explore CompactRuntime source
RELEASE MANIFEST RL-0013
88.5%BFCL v4 AST · full set
Parameters
27B dense
Adapter
LoRA r32 / α64
Architecture
262,144 tokens
RL sequence
Up to 65,536
License
Apache-2.0
88.5%BFCL v4 ASTOfficial full set · 1,240 cases
95.12%HumanEval156 / 164 · thinking mode
26.35%BigCodeBench-HardStrict pass@1 · 39 / 148
59.25%Functional testsSupplementary · 506 / 854

Tool use

The full BFCL
category profile.

The official 1,240-case BFCL v4 AST evaluation. This is the model's strongest verified public capability and the complete category table—not a selected subset.

Simple373 / 40093.2%
Multiple191 / 20095.5%
Parallel174 / 20087.0%
Parallel-multiple140 / 20070.0%
Irrelevance219 / 24091.2%
Overall1,097 / 1,24088.5%

What RL changed

A major repair.
Not a free lunch.

RL-0013 fixed the launch checkpoint's tendency to call tools when it should abstain. The category profile also shows the regression we do not hide.

Irrelevance43.8 → 91.2

The principal launch repair: the model learned when no tool call is relevant.

Overall BFCL82.7 → 88.5

Full-set AST accuracy after the RL stage.

Parallel-multiple93.0 → 70.0

Complex composition fell. RL changed the capability profile rather than improving every axis.

Deploy

Full quality.
Standard tooling.

Serve the pinned base plus adapter through vLLM, or load it with Transformers and PEFT. Thinking mode is recommended for coding.

vLLM
vllm serve Qwen/Qwen3.6-27B \
  --revision 6a9e13bd6fc8f0983b9b99948120bc37f49c13e9 \
  --served-model-name BTL-3 \
  --enable-lora --max-lora-rank 32 \
  --lora-modules BTL-3=/path/to/BTL-3 \
  --reasoning-parser qwen3 \
  --language-model-only --max-model-len 32768
Transformers + PEFT
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = "Qwen/Qwen3.6-27B"
revision = "6a9e13bd6fc8f0983b9b99948120bc37f49c13e9"
adapter = "badtheorylabs/BTL-3"

tokenizer = AutoTokenizer.from_pretrained(adapter)
model = AutoModelForCausalLM.from_pretrained(
    base, revision=revision, device_map="auto")
model = PeftModel.from_pretrained(model, adapter)

Local edition

Need the whole 27B model
in 8.39 GB?

BTL-3 Compact is a separate native artifact with its own runtime, retention profile, speed measurements, and deployment contract.

Open the Compact dossier ↗

Frozen identity

Verify the release.

adapter_model.safetensors · SHA-25637a8f519039707eba5906591cdb14268768db43f80489a9c2f83b3e51e5e89db933,974,032 bytes · Qwen3.6-27B revision 6a9e13bd6fc8f0983b9b99948120bc37f49c13e9

Reporting boundaries

LiveCodeBench: 88.08% is a completion-time-biased 193-of-442-case run, not a full leaderboard score.

HumanEval: saturated compatibility check; base Qwen3.6-27B scored 95.7% on the same harness.

Context: 262K is architectural. Launch benchmarks used 32K; RL sequences reached 65K.

Repository autonomy: no publishable SWE-bench or Terminal-Bench score is claimed.