GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
leinenterprises
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
leinenterprises
Model Tree
Qwen/Qwen3.8-27B
Input Modalities
Output Modalities
Supported Functionality
Options are labelled A, B, C, ... in the prompt. The answer is read from the label logits at the answer position and softmaxed over the listed options only. One forward pass per question, no sampling, no thinking.
A-Z, a-z, 0-9, then every two-letter pair AA..ZZ that is a single Qwen3.8 token (562 of them). Larger questions are refused as unsupported, never truncated. Questions with 62 or fewer options use only the single-character labels the model was trained on.max_pixels 4,000,000.bash
pip install vllm fastapi uvicorn peft transformers git+https://github.com/apolinario/decision-indexcd codepython -m train.merge --ckpt .. --out ../merged # adapter -> plain bf16 model (one-time)python -m serve.server --model ../merged --perms 1 --port 8000Then POST /v1/systemone with {"model", "state", "questions"}, the Decision Index wire format. state may be a string, a JSON object, or a list mixing text and {"image": "https://... | data:..."} parts. The code is in code/.
The same model also runs in-process as a kit engine: --engine serve.engine:JVEngine --option adapter=<path>.
| Benchmark | Result | Notes |
|---|---|---|
| typed-decisions (2,000 decisions) | acc 0.714, KL 0.318, Brier 0.129, ECE 0.026 | #7 overall; lowest ECE on the board |
| jev-bench test (22,773 records) | macro acc 0.734, ECE 0.084, Brier 0.334, TVD→human 0.391 | base in 4-bit NF4; Jev 1.13.0 API scores 0.733 / 0.113 / 0.349 / 0.432; files |
| Decision Index public text suite (140,178 rows) | index 49.58 | run |
| Path | median | mean | p80 |
|---|---|---|---|
| Text decisions, 1 client, A100 80GB, vLLM | 120 ms | 590 ms | 614 ms |
| Image decisions, RTX PRO 6000, vLLM | 254 ms | 399 ms | 635 ms |
The image row was measured with two option orders averaged; the shipped setting is one order. The full public 0.3 text suite (140,620 requests) took about 4 hours on one H100 with 16 concurrent clients.
Qwen/Qwen3.8-27B, LoRA r=32, alpha=64, on the language model's attention, linear-attention and MLP projections. The vision tower is frozen.SargeDev/jev-distill-corpus-v3.Benchmark numbers on this card come from our own replica of the public vision benchmarks, not the official board.
pip install vllm fastapi uvicorn peft transformers git+https://github.com/apolinario/decision-indexcd codepython -m train.merge --ckpt .. --out ../merged # adapter -> plain bf16 model (one-time)python -m serve.server --model ../merged --perms 1 --port 8000