GLM-5.2 is live. #1 throughput on OpenRouter, pay-per-token on FriendliAI. Try it today ➜
ProCreations
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
ProCreations
Model Tree
Qwen/Qwen3.8-27B
Input Modalities
Output Modalities
Supported Functionality
On one NVIDIA H200 with SGLang 0.5.17, native MTP measured 130.71 tok/s at temperature 0.7 and 125.42 tok/s at temperature 1.0, versus 69.27 and 69.26 tok/s for serial decode (1.89x and 1.81x). Eight single-stream samples were measured per temperature after warm-up with 256-token maximum completions.
For four 12k–13k-token prompt pairs, repeated-prefix reuse reduced median TTFT from
1.260 s to 0.522 s (2.41x). This cache result does not apply to a new arbitrary
prompt. See the raw JSON and server logs in benchmarks/ and the upstream model card
for capabilities, limitations, and recommended sampling settings.
A separate temperature-zero gate used SGLang's deterministic-inference mode with
FA3. Serial and native MTP generated identical text on 2 of 4 prompts, so this card
does not claim bitwise output identity for the accelerated path. The checkpoint is
still byte-identical to upstream; the mismatch is in numerical serving execution.
Use serial decode plus the engine's deterministic mode for strict output
reproducibility. The failed gate and both raw output sets are intentionally retained
in benchmarks/.