GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
MitakaKuma
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
MitakaKuma
Model Tree
Qwen/Qwen3.5-4B
Input Modalities
Output Modalities
Supported Functionality
Original paper results; checkpoint correspondence remains unverified. Scores: 0–100, higher is better.
BaT uses OpenHands; other systems use the Codex or Claude Code harness shown in the legend. Comparisons follow each benchmark's scorer; MedXpertQA-Text evaluates text responses.
Policy comparisons use the default OpenHands runner. Overall = (Task + Agentic) / 2. These differ from full-agent results above.
| Model / training | Overall | Agentic | Task |
|---|---|---|---|
| Qwen3.5-4B (base) | 6.1 | 12.1 | 0.0 |
| Qwen3.5-4B + SFT | 18.1 | 10.8 | 25.4 |
| Qwen3.5-4B + GRPO | 11.4 | 15.9 | 6.9 |
| BaT-4B (BiCuRL) | 22.9 | 28.8 | 17.0 |
xhigh effort, 64K output tokens per request, 200K context window.xhigh effort; Codex with 258K context and Claude Code with 1M context.Direct model API evaluation, compared only with the corresponding base model under the same protocol. Interactive tasks retain benchmark tools and environments.
| Benchmark | Qwen3.5-4B (base) | BaT-4B | Change (pp) |
|---|---|---|---|
| AIME 2025 | 85.0 | 56.7 | -28.3 |
| AIME 2026 | 89.7 | 53.3 | -36.4 |
| GPQA-Diamond | 76.2 | 58.1 |
bibtex
@article{liu2026bat, title = {BaT: Towards Self-Evolving Medical Research Agent with Stage Rubrics}, author = {Liu, Junqi and He, Yufan and He, Yexiao and Guo, Pengfei and Yang, Dong and Myronenko, Andriy and Zhao, Can and Ye, Hanrong and Qi, Tianhao and Zhou, Yuyin and Xu, Daguang and Tang, Yucheng}, journal = {arXiv preprint arXiv:2608.16211}, year = {2026}, url = {https://arxiv.org/abs/2608.16211}}| -18.1 |
| τ²-Bench | 79.9 | 29.1 | -50.8 |
| BFCL-Parity | 62.6 | 10.6 | -52.0 |
| GAIA | 15.2 | 2.4 | -12.8 |
| SWE-bench Verified | 39.4 | 4.8 | -34.6 |
| Terminal Bench 2.0 | 11.2 | 1.1 | -10.1 |