Use
Requires transformers>=5.2. Serve with vLLM and call it through the MIRA runtime:
vllm serve MIRA-MMDeepResearch/MIRA-SFT-35B-A3B --served-model-name MIRA \ --max-model-len 262144 --enable-auto-tool-choice --tool-call-parser qwen3_coder \ --allowed-local-media-path / --port 8000
python -m mira.harness.run --endpoint http://localhost:8000 --prompt-profile student \ --mode detail --question "..." --output runs/example
Sampling. temperature 1.0, top-p 0.95, top-k 20, presence penalty 1.5 (the values in
generation_config.json, except presence penalty, which the runtime sets).
Chat template. The bundled chat_template.jinja is the base model's template. Tools are described in the
system turn and called with the XML format <tool_call><function=NAME><parameter=KEY>VALUE</parameter></function></tool_call>.
Reasoning is on: keep enable_thinking true and pass preserve_thinking: true through
chat_template_kwargs so earlier reasoning stays in context across tool calls. The system prompt the model was
trained with is SYS_QWEN36_SFT_S2_MM in mira/agent/prompts/prompt.py; other system prompts reduce accuracy.
Evaluation
Accuracy (%) under the MIRA harness with the same judge for all agents. MMDeepResearch is the official
report score.
Table with columns: Model, Params, MMSearch, MMSearch+, BrowseComp-VL, SimpleVQA, LiveVQA, FVQA, VDR-Bench, MMDeepResearch| Model | Params | MMSearch | MMSearch+ | BrowseComp-VL | SimpleVQA | LiveVQA | FVQA | VDR-Bench | MMDeepResearch |
|---|
| OpenSearch-VL | 30B (3B) | 68.70 | 28.94 | 41.10 | 74.90 | 67.40 | 73.20 | 33.50 | 23.21 |
| Vision-DeepResearch | 30B (3B) | 69.60 | 28.50 | 53.70 | 66.33 | 77.60 | 74.20 | 37.80 | 12.78 |
| WeAgent-MMSearch | 30B (3B) | 77.78 | 37.82 | – | 77.73 | 81.05 | 79.33 | 35.68 | – |
| Qwen3.6-35B-A3B + MIRA-Harness | 35B (3B) | 73.10 | 33.44 | 56.64 | 74.33 | 78.33 | 70.00 | 19.00 | 41.90 |
| MIRA-SFT-35B-A3B | 35B (3B) | 80.70 | 38.91 | 58.40 | 79.00 | 82.00 | 79.33 | 29.80 | 42.61 |
"–" means not reported. Scores for other agents are from the MIRA paper.
Limitations
The report gate checks structure (length, images, sources, citations that resolve), not whether each claim is
true. Web content changes over time, so scores drift. The model was distilled from teacher trajectories and
inherits their errors.
Citation
@article{mira2027, title = {MIRA: A Multimodal Illustrated Deep Research Agent for Controllable Reports}, author = {Wu, Hongrui and Tuo, Kaiwen and Wang, Congcong and Li, Gen and Dong, Shuai and Yu, Xinlei and Hou, Haowen and Peng, Zelin and Wang, Jiaqi}, year = {2027}}