AgentMercury-Qwen3.5-35B-A3B is a Qwen3.5-35B-A3B (multimodal Mixture-of-Experts, Qwen3_5MoeForConditionalGeneration) 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 is the 35B-A3B member of the AgentMercury family; the smaller sibling is Minbyul/AgentMercury-Qwen3.5-4B.

Highlights

Base: Qwen3.5-35B-A3B (text + vision) — 35B total parameters, 3B activated per token, 256 experts (8 routed + 1 shared), 40 layers of hybrid Gated DeltaNet / gated attention, 262,144 native context.
Method: on-policy GRPO, fully-asynchronous MCP agentic RL (disjoint trainer / sglang rollout GPUs, 2 nodes × 8 GPUs).
Reward: final environment-state verifiers on real agent tasks (tool correctness + DB checks), with penalties for degeneration/truncation.
Rollout budget: global batch 128 (16 prompts × 8 samples), up to 20 tool-use turns and 24,576 response tokens per trajectory, constant LR 1e-6.
Training set: 38,670 create/update/delete agent tasks drawn from 3,950 distinct synthetic MCP environments.

Usage

Serving (sglang, recommended — matches evaluation)

bash

python3 -m sglang.launch_server \
  --model-path Minbyul/AgentMercury-Qwen3.5-35B-A3B \
  --served-model-name agentmercury-qwen3.5-35b-a3b \
  --host 0.0.0.0 --port 30000 --tp-size 1 \
  --context-length 131072 --mem-fraction-static 0.85 \
  --moe-runner-backend triton --attention-backend triton \
  --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).

Serve with at least a 128K context window. Like its base model, this checkpoint thinks before it answers; a short context window truncates generations inside the reasoning block and turns them into empty responses, which silently degrades any downstream score.

transformers

python

from transformers import AutoModelForCausalLM, AutoProcessor
model = AutoModelForCausalLM.from_pretrained(
    "Minbyul/AgentMercury-Qwen3.5-35B-A3B",
    torch_dtype="bfloat16", device_map="auto", trust_remote_code=True,
)
processor = AutoProcessor.from_pretrained(
    "Minbyul/AgentMercury-Qwen3.5-35B-A3B", trust_remote_code=True,
)

Training notes

These weights are the rollout-99 checkpoint of a 200-rollout MCP agentic RL run. Benchmark results for this checkpoint are not included in this card.

License

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