AgentMercury-Qwen3.5-4B is a Qwen3.5-4B (multimodal, Qwen3_5ForConditionalGeneration) checkpoint post-trained with agentic reinforcement learning on MCP (Model-Context-Protocol) tool-use environments. The RL objective rewards completing real multi-turn agent tasks (correct tool calls, correct final database/environment state), not just producing text.

This checkpoint is the clean-minimum of the run: the step at which reward peaks while degenerate-generation rate and truncation rate are both exactly 0 — before later steps regress into verbosity / reward-hacking.

Highlights

Base: Qwen3.5-4B (text + vision).
Method: on-policy GRPO, 200-step MCP agentic RL (slime trainer + sglang rollout, 8×A100).
Reward: final environment-state verifiers on real agent tasks (tool correctness + DB checks), with penalties for degeneration/truncation.
Training-set diversity: ~2.3k agent environments spanning 63% of industries and 76% of tools in the source corpus.

Benchmark results (improvement over the base model)

Only benchmarks where AgentMercury improves over the Qwen3.5-4B base are listed, with the absolute gain (Δ) over base. Evaluated with an OpenAI-compatible endpoint (sglang, 32k context), N repeats per cell.

Agentic / tool-use

Benchmark	Base	AgentMercury	Δ
BFCL	30.35	31.93	+1.58
τ³-bench	0.706	0.747	+0.041
τ²-bench	0.448	0.457	+0.009

Math & reasoning

Benchmark	Base	AgentMercury	Δ
AIME 2026	0.459	0.553	+0.094
HMMT 2026-02	0.285	0.356	+0.071
GPQA-Diamond	0.765	0.770	+0.005
Finance-Reasoning	0.563	0.571	+0.008
AA-Omniscience	−52.17	−51.67	+0.50

Code

Benchmark	Base	AgentMercury	Δ
LiveCodeBench (v5+v6)	0.366	0.435	+0.069
SciCode	0.226	0.260	+0.034

Writing

Benchmark	Base	AgentMercury	Δ
WritingBench	6.232	6.307	+0.075

Metrics use each benchmark's native scale (fractions 0–1, or the benchmark's own points). The largest, most consistent gains are on agentic tool-use (BFCL, τ-bench) and competition math / code (AIME, HMMT, LiveCodeBench).

Usage

Serving (sglang, recommended — matches evaluation)

bash

python3 -m sglang.launch_server \
  --model-path Minbyul/AgentMercury-Qwen3.5-4B \
  --served-model-name agentmercury-qwen3.5-4b \
  --host 0.0.0.0 --port 30000 --tp 1 \
  --context-length 32768 \
  --reasoning-parser qwen3 --tool-call-parser qwen3_coder \
  --trust-remote-code

Then call the OpenAI-compatible endpoint at http://localhost:30000/v1 (supports tool calls).

transformers

python

from transformers import AutoModelForCausalLM, AutoProcessor
model = AutoModelForCausalLM.from_pretrained(
    "Minbyul/AgentMercury-Qwen3.5-4B",
    torch_dtype="bfloat16", device_map="auto", trust_remote_code=True,
)
processor = AutoProcessor.from_pretrained(
    "Minbyul/AgentMercury-Qwen3.5-4B", trust_remote_code=True,
)

Training notes

The reward peaks around this checkpoint while the model stays clean (no repetition collapse, no context truncation). Continuing RL past this point raised response length and reintroduced degeneration/truncation without adding capability — so this clean-minimum checkpoint is released as the recommended weights.

License

Released under the Apache-2.0 license (see LICENSE).