GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
baseten
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
baseten
Model Tree
Qwen/Qwen3.8-27B
Input Modalities
Output Modalities
Supported Functionality
| Checkpoint | Accuracy | Correct / scored completions | Truncation |
|---|---|---|---|
| Pinned base model, before environment training | 1.25% | 5 / 400 | 5.25% |
| Final expert, update 200 | 27.50% | 110 / 400 | 9.75% |
Accuracy improved by 26.25 percentage points from the pinned base model. Each evaluation uses 50 held-out tasks × 8 sampled completions = 400 scored completions. Accuracy is the mean binary correctness across all completions, not best-of-eight or pass-at-eight. Truncated responses remain in the denominator and count as failures. The reported endpoint is the final checkpoint, not an earlier peak.
The released Figlet expert is the final update-200 checkpoint. The first 100 updates ended at 22.25% accuracy and 18.50% truncation. Training then continued from the full checkpoint, restoring model, Adam, scheduler, RNG, and data-cursor state for 100 additional updates. The added training improved final accuracy by 5.25 percentage points from that update-100 endpoint. An intermediate update-150 evaluation reached 33.75%; it is not the released checkpoint.
The bank contains 400 training cases and 50 separate held-out validation cases. Validation uses a dedicated inference role every 10 optimizer updates with fixed seeds [17, 43, 101, 137, 211, 307, 401, 503]. All scheduled evaluations and final inference/full-state archival were verified against native receipts. The same validation bank was inspected during training; there was no separate final confirmation set. Results establish performance on held-out cases within this environment and do not establish cross-environment transfer or broad general reasoning improvement.
| Setting | Value |
|---|---|
| Training | Full-weight GRPO; Adam |
| Learning rate | Constant 1e-6 |
| Adam betas / weight decay | (0.9, 0.98) / 0.1 |
| Policy clip epsilon / upper epsilon | 0.2 / 0.28 |
| KL / entropy coefficients | 0.0 / 0.0 |
| Batch per optimizer update | 8 questions × 8 completions = 64 responses |
| Total consumed training responses | 12,800 |
| Precision | BF16 model training and sampling |
| Thinking | Enabled; reasoning_effort="medium" |
| Sampling | Temperature 1.0, top-p 0.95, top-k 20 |
| Maximum generated tokens | 16384 |
| Context limit used in this experiment | 32768 |
| Checkpoint / validation interval | Every 10 optimizer updates |
| Async consumption | Ordered batches; maximum completed-update lag of 1 |
For comparable text generation, use the supplied tokenizer and chat template, keep thinking enabled with reasoning_effort="medium", and retain the sampling and output limits above. The original runtime pinned Transformers 5.12.1; detailed runtime revisions are recorded in the provenance JSON. This expert was trained and evaluated on text inputs. Vision or video performance and retention of unrelated base-model capabilities were not evaluated.
The source inference snapshot is miles-qwen3.8-27b-009f382af670/extend-977cb492bfef/training/eval-snapshots/step_199. It was emitted at optimizer update 200 and verified complete in durable platform storage. The native final evaluation completed at 2026-10-19T23:38:33.645206+00:00. This release contains inference weights and required inference assets; resumable optimizer state and raw training/evaluation traces remain in the experiment archives.
W&B's final validation point is missing: its recorded validation history stops at update 190 (29.25%). An existing-run update permission restriction prevents reconciliation. The final 27.50% at update 200 reported here comes from verified native evaluation receipts and archived rollout evidence; the W&B gap does not change that measured endpoint.
The legacy adapter recorded some detailed non-truncation categories as unknown despite valid binary rewards. Detailed grader labels were recovered into a separate derived report from saved grader reasons; original rewards and raw outcomes were preserved. Accuracy above uses the scored binary rewards.
This is a specialized research checkpoint for text reasoning in the named environment. Its results depend on the frozen bank, prompt and grading contract, decoding settings, and token budget. Broader task transfer, production safety, multimodal behavior, and performance under different decoding settings have not been established by these experiments.
Apache-2.0, following the pinned base model. See the accompanying license and the pinned base-model card.
Both experts below completed 200 optimizer updates. Adam used K=8 and 12,800 consumed training completions; SGD used K=16 and 25,600. Matching update counts does not isolate the optimizer's effect.
Training plots retain all 200 recorded raw-reward points on the original rollout/step axis 0–199, corresponding to the batches consumed by updates 1–200. The displayed running average uses a window of 10 with the original series visible and full-fidelity data. Validation is unsmoothed and includes the baseline plus every ten completed updates through 200; the final points come from verified native evaluation receipts where the original scalar histories omit them. The images show W&B-rendered plots of the recorded data; no training, evaluation or rescore was performed to produce them.
Chart provenance, raw values and settings. Raw outcome categories are retained separately from mean binary rewards; these images do not imply a new correctness-count reconstruction.