GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
fcreait
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
fcreait
Model Tree
Qwen/Qwen3.5-4B
Input Modalities
Output Modalities
Supported Functionality
model.safetensors-* — the base model with the LoRA merged in, stored in float16. The
merge is done in float32 and saved in float16 on purpose: saving in bfloat16 rounds away
a large part of the fine-tune's update. The GGUF files were converted from exactly these
weights.model-mtp.safetensors — Qwen's own multi-token-prediction head from the base model,
unchanged apart from the float16 cast. It is not fine-tuned.lora/ — the LoRA adapter itself (r16, α32, checkpoint 200 of 413), for anyone who
prefers to apply it to the base model.Everything else (tokenizer, chat template, vision processor config) is the base model's.
Load it as you would Qwen3.5-4B. Turn thinking off (enable_thinking: false): the
fine-tune was trained and evaluated without it. The base model's recommended non-thinking
sampling (temperature 0.7, top-p 0.8, top-k 20, presence penalty 1.5) is what we tested.
Apache-2.0, the same as the base model.