TIR-Qwen3.5-9B-NonThinking-v2 API & Inference Endpoint

Training data

Table with columns: Component, Rows
Component	Rows
AgenticData true-multiturn SFT (extreme-context filtered)	4,292
CodeDance SFT (multi-turn `<code>` tool use)	13,828
Total	18,120

No assistant message contains a <think> block; the empty non-thinking prefix is inserted by the template and excluded from the loss.

Training configuration

Base model: Qwen/Qwen3.5-9B; full-parameter SFT, vision tower and aligner frozen
Recipe: non-thinking (add_non_thinking_prefix=true, loss_scale=default+ignore_empty_think, truncation_strategy=delete)
Learning rate 1e-5, min 1e-6, cosine decay, warmup fraction 0.05, weight decay 0.1, clip grad 1.0
Maximum length 32,768; global batch 64; micro batch 1; TP=2, PP=1, sequence parallel; bf16; flash attention
One epoch = 283/283 optimizer steps, elapsed 8h 10m 55s, final training loss 0.17167911
Full arguments: training/args.json; per-step log: training/logging.jsonl

Evaluation

Not yet evaluated at publication time. Results on TIR-Bench and the huawei_benchmark suite (FVQA, SimpleVQA, MMBrowseComp, GAIA) will be added alongside the legacy checkpoint for comparison.

Inference

bash
vllm serve HEART77/TIR-Qwen3.5-9B-NonThinking-v2 --port 8000 --max-model-len 32768 --reasoning-parser qwen3

Disable thinking at request time with chat_template_kwargs={"enable_thinking": false}. The model uses the TIR agent format: Python in <code> blocks, tool results as a user turn beginning with Code execution result:, final answer in <answer>...</answer>.

Training data

Table with columns: Component, Rows
Component	Rows
AgenticData true-multiturn SFT (extreme-context filtered)	4,292
CodeDance SFT (multi-turn `<code>` tool use)	13,828
Total	18,120

No assistant message contains a <think> block; the empty non-thinking prefix is inserted by the template and excluded from the loss.

Training configuration

Base model: Qwen/Qwen3.5-9B; full-parameter SFT, vision tower and aligner frozen
Recipe: non-thinking (add_non_thinking_prefix=true, loss_scale=default+ignore_empty_think, truncation_strategy=delete)
Learning rate 1e-5, min 1e-6, cosine decay, warmup fraction 0.05, weight decay 0.1, clip grad 1.0
Maximum length 32,768; global batch 64; micro batch 1; TP=2, PP=1, sequence parallel; bf16; flash attention
One epoch = 283/283 optimizer steps, elapsed 8h 10m 55s, final training loss 0.17167911
Full arguments: training/args.json; per-step log: training/logging.jsonl

Evaluation

Not yet evaluated at publication time. Results on TIR-Bench and the huawei_benchmark suite (FVQA, SimpleVQA, MMBrowseComp, GAIA) will be added alongside the legacy checkpoint for comparison.

Inference

bash
vllm serve HEART77/TIR-Qwen3.5-9B-NonThinking-v2 --port 8000 --max-model-len 32768 --reasoning-parser qwen3

TIR-Qwen3.5-9B-NonThinking-v2

README

Training data

Training configuration

Evaluation

Inference

Explore FriendliAI today

README

Training data

Training configuration

Evaluation

Inference