Qwen3.5-9B-Distill-Agent-Instruct

Qwen3.5-9B-Distill-Agent-Instruct is an agent- and instruction-oriented distilled post-training of Qwen3.5-9B. It learns from a curated mixture of multi-model agent trajectories and is tuned for reliable tool-use decisions, structured function calling, multi-step agent execution, code-oriented tasks, and direct instruction following.

This repository provides the BF16 merged model weights. The trained rsLoRA adapter has already been merged into the base model, so no separate PEFT adapter is required for inference.

Highlights

• Agent-oriented post-training: trained on long-horizon interaction traces containing planning, tool calls, tool responses, intermediate state, recovery behavior, and final answers.
• Stronger tool-use boundaries: explicitly learns both when a tool is necessary and when a direct answer is preferable.
• Multiple and parallel function calling: improves coordination across several calls and concurrent tool execution patterns.
• Irrelevance handling: better resistance to invoking tools that do not match the user's request.
• Structured output: trained on explicit function schemas and tool-call serialization for easier integration with agent runtimes.
• Code and instruction following: the mixture includes cleaned code examples and general instruction trajectories alongside agent data.
• Long-context foundation: inherits Qwen3.5-9B's native 262,144-token context window.
• Multimodal architecture retained: the vision tower and aligner were frozen during this post-training stage; the reported gains focus on language-side Agent and Tool Call behavior.

BFCL Tool-Call Evaluation

The model was evaluated on BFCL (Berkeley Function Calling Leaderboard) after LoRA post-training. The table below reports the supplied full-run results. Base is the original Qwen3.5-9B checkpoint and Distill Agent Instruct is the merged post-trained model. Δ is the absolute change in percentage points.

BFCL category	Samples	Base	Distill Agent Instruct	Δ
Simple Python	400	92.00%	91.50%	-0.50 pp
Multiple	200	94.00%	96.00%	+2.00 pp
Parallel	200	85.50%	90.50%	+5.00 pp
Parallel Multiple	200	87.00%	86.00%	-1.00 pp
Irrelevance	240	83.75%	89.17%	+5.42 pp
BFCL Non-Live Overall	—	74.29%	75.75%	+1.46 pp

The most substantial improvements appear in parallel function calling and irrelevance detection. This is consistent with the training mixture's emphasis on multi-action trajectories and balanced call-versus-answer decisions. Simple Python and parallel-multiple cases show small regressions, so the result should be read as a capability trade-off rather than uniform improvement across every BFCL category.

Latency in the reported evaluation

Model	Mean latency	P95 latency
Qwen3.5-9B Base	2.96 s	6.07 s
Qwen3.5-9B-Distill-Agent-Instruct	2.54 s	5.60 s

Under the supplied test setup, mean latency decreased by approximately 14.2%, while P95 latency decreased by approximately 7.7%. Latency depends on serving engine, hardware, batching, generation parameters, and output length; these figures should only be compared within the same evaluation setup.

Training Dataset

Post-training used Fable5 · GPT-5.5 · Opus-4.7 Mixed Agent Traces, an author-curated mixture containing:

• Claude Fable5 long-horizon agent trajectories;
• GPT-5.5 Agent, GLM-5.2 Agent, and Qwen3.7-Max PI traces;
• LiteLLM step-level trajectories;
• Claude Opus trace-inversion examples;
• Glaive and Hermes function-calling conversations;
• rebalanced when-to-call supervision;
• cleaned code instruction data.

The public release contains 20,409 records, 211,580 message events, and 67,865 explicit tool-call steps. Its dedicated processing pipeline includes schema normalization, trajectory-integrity checks, meaningful step slicing, judge-gated filtering, source-aware sampling, structural/content deduplication, and tool-use rebalancing.

LoRA Post-Training Configuration

Training used rank-stabilized LoRA (rsLoRA) through PEFT. The adapter targeted all linear layers in model.language_model, while the vision tower and multimodal aligner remained frozen.

Parameter	Value
Method	PEFT LoRA with rsLoRA enabled
Rank (r)	64
Alpha	128
Dropout	0.05
Bias	None
Target scope	All language-model linear layers
Target projections	q_proj, k_proj, v_proj, o_proj, out_proj, in_proj_a, in_proj_b, in_proj_z, in_proj_qkv, gate_proj, up_proj, down_proj
DoRA / QA-LoRA	Disabled / Disabled
Vision tower / aligner	Frozen / Frozen
Training and merge precision	BF16

Optimization setup

Parameter	Value
Epochs	3
Maximum sequence length	32,768 tokens
Learning rate	1e-4
Scheduler	Cosine
Warmup ratio	0.1
Weight decay	0.1
Per-device micro-batch size	3
Gradient accumulation	2
Maximum gradient norm	1.0
Optimizer	Fused AdamW
Runtime	BF16, gradient checkpointing, Liger Kernel, DeepSpeed ZeRO-3

Base Architecture

Qwen3.5-9B-Distill-Agent-Instruct retains the compact hybrid vision-language architecture of Qwen3.5-9B:

• 9B language-model parameters, 32 layers, and hidden dimension 4096;
• hybrid Gated DeltaNet and Gated Attention blocks;
• native 262,144-token context window;
• Qwen3.5 tokenizer, chat template, and multimodal input stack;
• Multi-Token Prediction support inherited from the base checkpoint.

For full architectural details and base-model benchmark results, refer to the Qwen3.5-9B model card.

Serving

The model can be served directly with an OpenAI-compatible endpoint. Use a recent vLLM build with Qwen3.5 support:

vllm serve lzy510016411/qwen3.5-9b-distill-agent-instruct \
  --port 8000 \
  --tensor-parallel-size 1 \
  --max-model-len 262144 \
  --reasoning-parser qwen3 \
  --enable-auto-tool-choice \
  --tool-call-parser qwen3_coder

If memory is limited, reduce --max-model-len. For text-only workloads, add --language-model-only to skip vision profiling and reserve more memory for KV cache.

The included chat template supports normal assistant responses and structured tool calls. Applications should pass tool definitions through the serving framework's OpenAI-compatible tools field and keep tool-call/tool-response ordering intact across turns.

Intended Use

This checkpoint is intended for research and application development involving instruction following, agent orchestration, function calling, tool selection, multi-step task execution, and coding assistance.

Tool execution should remain sandboxed and subject to application-level authorization. BFCL results measure structured function-calling behavior and do not establish factual correctness, safety, or suitability for autonomous high-stakes actions.

License and Attribution

This derivative checkpoint follows the base model's Apache 2.0 license. See the Qwen3.5-9B license for details.

@misc{qwen3.5,
  title  = {{Qwen3.5}: Towards Native Multimodal Agents},
  author = {{Qwen Team}},
  month  = {February},
  year   = {2026},
  url    = {https://qwen.ai/blog?id=qwen3.5}
}