gl-agent-1-27b

gl-agent-1-27b

27B parameters, fine-tuned for long horizon agentic tasks.

Serving

vllm serve <path> --served-model-name gl-agent-1-27b \
  --max-model-len 262144 \
  --enable-auto-tool-choice --tool-call-parser qwen3_xml \
  --reasoning-parser qwen3 \
  --speculative-config '{"method":"mtp","num_speculative_tokens":1}'

Context is 262,144 natively and extensible to 1,000,000 tokens with YaRN. Lower --max-model-len if your hardware can't hold the full window.

Tool calling

Tool calls are emitted as XML, so --enable-auto-tool-choice and --tool-call-parser qwen3_xml are both required. Without them tool_calls comes back null and the raw markup lands in content.

from openai import OpenAI
client = OpenAI(base_url="http://localhost:8000/v1", api_key="none")

tools = [{"type": "function", "function": {
    "name": "run_bash",
    "description": "Run a shell command",
    "parameters": {"type": "object",
                   "properties": {"cmd": {"type": "string"}},
                   "required": ["cmd"]}}}]

r = client.chat.completions.create(
    model="gl-agent-1-27b", tools=tools, temperature=1.0, top_p=0.95,
    messages=[{"role": "user",
               "content": "The test suite fails. Find out why, start with the CI config."}])

print(r.choices[0].message.tool_calls[0].function)
# Function(arguments='{"cmd": "ls -la . && find . -name \\"*.yml\\" ... | head -20"}',
#          name='run_bash')

Feed results back as {"role": "tool", "name": ..., "content": ...} and the model continues the loop.

MTP

The multi-token-prediction head ships inside the weights, so speculative decoding needs no separate draft model. Pass it as shown above.

Measured mean acceptance length is about 1.7 against a ceiling of 2.0 at num_speculative_tokens: 1, so the drafted token is taken roughly 70% of the time. The head is one layer deep, which makes 1 its natural setting. Higher values are accepted but re-run the same layer and acceptance falls off. How much wall-clock that buys depends on your memory bandwidth, so measure it on your own hardware rather than trusting a number from someone else's.

Thinking

On by default. Sampling defaults ship in generation_config.json, and enable_thinking, preserve_thinking and reasoning_effort are settable per request through chat_template_kwargs:

extra_body={"chat_template_kwargs": {"enable_thinking": False}}

Benchmarks

Coming soon.