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
ProCreations/Qwen-3.8-SuperFast
Input Modalities
Output Modalities
Supported Functionality
The native one-layer MTP proposal was initialized from the base checkpoint and trained with a frozen target. Its objective combines:
For a single speculative token, 1 - TV(target, draft) is the ideal acceptance
probability before implementation overhead. That makes the TV term directly
connected to the intended speed mechanism. The training objective uses full
vocabulary distributions, not a top-k distillation shortcut.
The best checkpoint was selected on held-out conversations. The training report
is stored at tempmtp/training-report.json, and the standalone native head is at
tempmtp/mtp-head.safetensors.
Use the same native Qwen3.5 MTP path as the base release:
bash
qwen-superfast plan \--variant bf16 \--method mtp \--model ProCreations/Qwen-3.8-SuperFast-TempMTP
The target verifier is still the unchanged BF16 model. This checkpoint is a speed experiment, not a claim that speculative execution is bitwise identical to serial execution. Use target-only serial decoding when bitwise execution is required.
The training-time held-out acceptance estimate improved. End-to-end SGLang throughput is pending until the matched benchmark artifact is published.
Temperature-specific throughput must be measured on the same GPU, engine build, prompts, sampling parameters, and output length as the native-MTP baseline.
ProCreations/Qwen-3.8-SuperFast@196d97c3c5533e3e36e442ca8a1cfaa987692d81HuggingFaceH4/ultrachat_200k (train_sft)mtp.* tensorsThis model inherits the base model's Apache-2.0 license and usage limitations.