BTL-3

A 27B open-weight agent model for agentic coding and structural tool use

95.1% HumanEval · 88.5% BFCL v4 AST · 88.1% LiveCodeBench v6 (193-case run)

Compact edition · Runtime source · Bad Theory Labs · Discord

Introducing BTL-3

BTL-3 is a 27B open-weight agent model built for agentic coding, structural tool use, repository work, failure recovery, and long multi-turn execution.

The release includes the full BTL-3 model and BTL-3 Compact, which packages the complete text model into one 8.39 GB native file. That is smaller than an 8B model stored in FP16 and corresponds to an effective artifact footprint of under 2.5 bits per parameter. On a fresh private 100-turn tool-contract gate, BTL-3 Compact retained 83 of the 90 behaviors the full model completed correctly: 92.2% measured conditional tool-behavior retention.

Overview

BTL-3 is a post-trained Qwen3.6-27B model for coding agents, repository work, structured tool use, and long multi-turn execution. It is tuned to reason, act, inspect tool results, recover from failures, and stop when no action is required.

This repository contains the frozen RL-0013 rank-32 PEFT adapter and its tokenizer configuration. The base checkpoint is pinned to an exact revision for reproducible loading.

Highlights

• Strong structured tool use across single, multiple, and parallel calls.
• 91.2% BFCL irrelevance, measuring when the model correctly avoids an unnecessary tool call.
• Thinking-mode coding with 95.12% HumanEval pass@1.
• Qwen3.6 hybrid-attention architecture with a declared 262,144-token context window.
• Open weights under Apache-2.0, deployable with Transformers or vLLM.
• An independent 8.39 GB Compact edition is available for native local inference.

Results

All values below belong to the frozen BTL-3 RL-0013 release.

Evaluation	Score	Protocol
BFCL v4 AST	88.5% (1097/1240)	Complete official full set
HumanEval	95.12% (156/164)	pass@1, thinking mode
LiveCodeBench v6	88.1% (170/193)	Completed 193-case run, thinking mode
BigCodeBench-Hard Instruct	26.35% (39/148)	Official strict pass@1
BigCodeBench functional tests	59.25% (506/854)	Supplementary test-level score

BFCL v4 category breakdown

Category	Score
Simple	93.2%
Multiple	95.5%
Parallel	87.0%
Parallel-multiple	70.0%
Irrelevance	91.2%

Model specification

Item	Specification
Base model	Qwen/Qwen3.6-27B
Base revision	6a9e13bd6fc8f0983b9b99948120bc37f49c13e9
Release checkpoint	BTL-3 RL-0013
Adapter	PEFT LoRA, rank 32, alpha 64
Adapter size	933,974,032 bytes
Architectural context	262,144 tokens
Maximum RL sequence length	65,536 tokens
Launch benchmark context	32,768 tokens
Recommended mode	Thinking enabled for coding and reasoning
License	Apache-2.0

Quickstart

Transformers

import torch
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base_id = "Qwen/Qwen3.6-27B"
base_revision = "6a9e13bd6fc8f0983b9b99948120bc37f49c13e9"
adapter_id = "badtheorylabs/BTL-3"

tokenizer = AutoTokenizer.from_pretrained(adapter_id)
base = AutoModelForCausalLM.from_pretrained(
    base_id,
    revision=base_revision,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)
model = PeftModel.from_pretrained(base, adapter_id)

messages = [
    {
        "role": "user",
        "content": "Inspect this repository, fix the failing tests, and explain the patch.",
    }
]
prompt = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
    enable_thinking=False,
)
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
output = model.generate(**inputs, max_new_tokens=4096)
completion = output[0, inputs.input_ids.shape[1]:]
print(tokenizer.decode(completion, skip_special_tokens=False))

The supported release default is non-thinking mode. Thinking remains available for controlled evaluation, but it is currently discouraged because the RL-0013 reasoning policy can become repetitive or fail to terminate on some prompts.

vLLM

BTL-3 was evaluated with vLLM 0.23.0:

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 \
  --lora-target-modules \
    q_proj k_proj v_proj o_proj \
    in_proj_qkv in_proj_z in_proj_b in_proj_a out_proj \
    gate_proj up_proj down_proj \
  --reasoning-parser qwen3 \
  --language-model-only \
  --max-model-len 32768

For structured tools, enable the Qwen XML tool parser supported by your installed vLLM version.

Intended use

• coding, debugging, and test-driven repair;
• repository and terminal agents;
• structured function calling and multi-tool workflows;
• private or self-hosted agent deployments;
• long multi-turn tasks that require verification and recovery.

Artifact integrity

Artifact	SHA-256
adapter_model.safetensors	37a8f519039707eba5906591cdb14268768db43f80489a9c2f83b3e51e5e89db

Operational guidance

Run generated code and tool calls in a sandbox. Require explicit confirmation before destructive, privileged, financial, or otherwise high-impact actions.

License and citation

The adapter is released under Apache-2.0 and requires the separately distributed Qwen3.6-27B base model.

@software{btl3_2026,
  title  = {BTL-3: A 27B Agentic Coding and Tool-Use Model},
  author = {Bad Theory Labs},
  year   = {2026},
  url    = {https://huggingface.co/badtheorylabs/BTL-3}
}

For questions and release updates, visit Bad Theory Labs or join the community Discord.