GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
ZACK0225
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
ZACK0225
Model Tree
Qwen/Qwen3.8-27B
Input Modalities
Output Modalities
Supported Functionality
1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0.claude-opus-5; one preselected replicate, taking the first valid outcome after service-error retries.Detailed settings and package versions are in training_recipe.json and deepspeed-zero3.json.
| Revision | Epoch | Steps | MAB validation loss |
|---|---|---|---|
epoch-1 | 1 | 141 | 0.2227127254 |
main | 2 | 282 | 0.2212697268 |
The loss is the mean of per-turn assistant-token mean cross-entropies, not a benchmark accuracy. The 275 external validation tasks and 275 final-test tasks are excluded from training; final-test results are excluded from checkpoint selection. Full MAB evaluation will distinguish training tasks from tasks not used for training. Evaluation results will be documented when complete.
The export retains the Qwen3_5ForConditionalGeneration architecture used by the Qwen3.8 base model. It includes full weights, tokenizer, chat template, and processor configuration. No adapter merge is needed.
python
import torchfrom transformers import AutoTokenizer, Qwen3_5ForConditionalGeneration repo_id = "ZACK0225/Qwen3.8-27B-MAB-Opus5-SFT"revision = "main"tokenizer = AutoTokenizer.from_pretrained(repo_id, revision=revision)model = Qwen3_5ForConditionalGeneration.from_pretrained( repo_id, revision=revision, dtype=torch.bfloat16, device_map="auto")messages = [{"role": "user", "content": "Explain exploration and exploitation in a bandit task."}]inputs = tokenizer.apply_chat_template( messages, tokenize=True, add_generation_prompt=True, return_tensors="pt").to(model.device)with torch.inference_mode(): output = model.generate(inputs, max_new_tokens=256)print(tokenizer.decode(output[0, inputs.shape[1]:], skip_special_tokens=True))Training used Transformers 5.20.2. Local serving was exercised with vLLM 0.20.2. The complete BF16 weights occupy approximately 54.7 GB; inference additionally requires memory for runtime state and the KV cache. Multimodal capability is inherited from the base model; this SFT used text trajectories and did not evaluate vision transfer.
The upstream Apache-2.0 license is included as LICENSE; modifications are documented in NOTICE. File hashes are listed in SHA256SUMS.
import torchfrom transformers import AutoTokenizer, Qwen3_5ForConditionalGeneration repo_id = "ZACK0225/Qwen3.8-27B-MAB-Opus5-SFT"revision = "main"tokenizer = AutoTokenizer.from_pretrained(repo_id, revision=revision)model = Qwen3_5ForConditionalGeneration.from_pretrained( repo_id, revision=revision, dtype=torch.bfloat16, device_map="auto")messages = [{"role": "user", "content": "Explain exploration and exploitation in a bandit task."}]inputs = tokenizer.apply_chat_template( messages, tokenize=True, add_generation_prompt=True, return_tensors="pt").to(model.device)with torch.inference_mode(): output = model.generate(inputs, max_new_tokens=256)print(tokenizer.decode(output[0, inputs.shape[1]:], skip_special_tokens=True))