Qwen3.6 27B

fp8_channel_dynamic_s2 quantized and serving-tuned for H100 on vLLM 0.25.1, with the benchmark receipt included. You get a self-contained kit: the optimized weights, the exact serving configuration, and the measured proof. Run it on any cloud or your own hardware. Nothing calls home, nothing expires.

$40

one-time, yours forever

Get this model

Speedup: 1.2xfaster p50
Latency p50: 2215ms
Latency p99: 2296ms
Peak VRAM: 71.0GB
GPU target: H100
License: Apache-2.0

Latency p50-22%

Latency p99-20%

Peak VRAM0%

Throughput+29%

Measured against the unoptimized baseline on H100: median latency 2857 ms down to 2215 ms. The full benchmark methodology and receipt ship inside the kit.

This is what it means to own your AI

You control the intelligence

The weights run on your infrastructure. No deprecation notices, no silent model swaps, no provider deciding what your product is allowed to do.

You keep the economics

One purchase instead of a meter that runs forever. A closed API charges you for every token, at whatever the price becomes; your own deployment costs what your GPU costs.

You keep the data

Prompts and outputs never leave your network. Nothing calls home, nothing is retained by someone else, and compliance stops being a negotiation.

Technical implementation

Optimizations applied

Channelwise FP8 on MLP projections only
attention, Gated-DeltaNet, vision tower, lm_head all at BF16
Maximum sensitive-layer protection
the most conservative quantized tier; chosen by measured accuracy ladder
Pinned engine + measured serving configuration
identical harness as the max-speed tier; MTP disabled by measurement

Serving stack

vLLM 0.25.1, fp8_channel_dynamic_s2, tuned for H100 (sm_90). Requires 80 GB VRAM minimum.

License

Apache-2.0, as published by the model author on Hugging Face. The full license text ships inside the kit.

Runs on any GPU provider

The kit is self-contained: bring it to any cloud with H100-class GPUs, or your own hardware.

Docker-composeKubernetesRunpodModalYour own hardware