GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
Guanghan-Ning
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
Guanghan-Ning
Model Tree
Qwen/Qwen3.8-27B
Input Modalities
Output Modalities
Supported Functionality
| Base model | Qwen3.8-27B (dense) |
| Algorithm | GRPO, one update per rollout step; the advantage of each episode is given to every call of that episode ("broadcast" credit) |
| Reward | outcome only: +1 per level solved, +0.5 for clearing all levels, −0.005 per model call |
| Data | 22 training games, 8 episodes per game per step (global batch 176 episodes) |
| Steps | 20 (this checkpoint is the state after the 20th update) |
| Harness | WitnessBench harness v1 (the agent harness of the paper), thinking off; text only. The current release ships harness v2; this checkpoint was not trained with it |
| Optimizer | Adam (0.9, 0.98), lr 1e-6 constant, weight decay 0.1, grad clip 1.0, clip ε 0.2/0.28, no KL term |
| Rollout | temperature 1.0, top-p 1.0, up to 8,192 tokens per call |
| Hardware | 8 × NVIDIA B300, miles trainer |
The model is used with thinking off (enable_thinking: false in the chat template), which is the dialect it was
trained in, except where an evaluation below says otherwise. Example with SGLang:
bash
python3 -m sglang.launch_server --model-path Guanghan-Ning/WitnessGym-Qwen3.8-27B-RL-broadcast-step20 \ --tp 1 --context-length 32768 --sampling-defaults openai \ --default-chat-template-kwargs '{"enable_thinking": false}'--sampling-defaults openai samples with temperature 1.0, top-p 1.0 and no top-k unless a request sets them, which is
how every evaluation of this checkpoint was run (the shipped generation_config.json is the base model's).
Out-of-domain transfer (paper; mean ± std over repetitions, paired difference to the base model with a 95% bootstrap interval):
| Benchmark (metric) | Setting | Base | This model | Δ |
|---|---|---|---|---|
| CipherBank (custom ciphers, exact match %) | thinking on, 32k, 2 reps | 73.8 ± 0.2 | 78.6 ± 1.3 | +4.8 [+1.5, +8.1] |
| PhysGym L4 (success %) | thinking on, 32k, 3 reps | 44.7 ± 1.6 | 48.1 ± 2.1 | +3.4 [0.0, +7.2] |
| WILT (elim@5 %) | thinking off, 5 reps | 86.3 ± 1.6 | 88.6 ± 1.7 | +2.3 [+0.3, +4.4] |
| Failing-to-Falsify (incompatibility fraction %) | thinking off, 5 reps | 50.7 ± 5.1 | 55.4 ± 4.9 | +6.1 [−0.2, +12.2] |
How to reproduce these numbers: see transfer/ in the WitnessBench repository.
Apache 2.0, the license of the base model.
bibtex
@article{ning2026witness, title = {Witness: Discovery, Deciphering, and Epiphany in Interactive Puzzle Environments}, author = {Ning, Guanghan and Liu, Ping and Li, Linyi and Zheng, Huangjie and Neervannan, Arjun and Nguyen, Huu and Sklar, Michael and Zorlu, Deniz and Ouporov, Nicolai}, journal = {arXiv preprint arXiv:2609.32208}, year = {2026}, eprint = {2609.32208}, archivePrefix = {arXiv}, primaryClass = {cs.AI}, doi = {10.48550/arXiv.2609.32208}, url = {https://arxiv.org/abs/2609.32208}}python3 -m sglang.launch_server --model-path Guanghan-Ning/WitnessGym-Qwen3.8-27B-RL-broadcast-step20 \ --tp 1 --context-length 32768 --sampling-defaults openai \ --default-chat-template-kwargs '{"enable_thinking": false}'@article{ning2026witness, title = {Witness: Discovery, Deciphering, and Epiphany in Interactive Puzzle Environments}, author = {Ning, Guanghan and Liu, Ping and Li, Linyi and Zheng, Huangjie and Neervannan, Arjun and Nguyen, Huu and Sklar, Michael and Zorlu, Deniz and Ouporov, Nicolai}, journal = {arXiv preprint arXiv:2609.32208}, year = {2026}, eprint = {2609.32208}, archivePrefix = {arXiv}, primaryClass = {cs.AI}, doi = {10.48550/arXiv.2609.32208}, url = {https://arxiv.org/abs/2609.32208}}