Qwen3.8-27B, Mirai S, experimental

Qwen3.8-27B quantized with the Mirai S codec: 2.4 bits per weight, 8.45 GB on disk.

This is a work in progress and the numbers here are not final. Work on both runtimes is unfinished. Treat anything measured with it as provisional.

On Apple silicon it runs on a prebuilt uzu runtime that ships in this repo. On NVIDIA it runs on stock vLLM with a small plugin that also ships in this repo.

Apple silicon — uzu

You need an Apple silicon Mac on macOS 26.4 or newer with 24 GB of memory or more. It runs on every Apple silicon chip and is fastest on M5, whose matrix units it uses. M1 to M4 run it correctly but more slowly: about 4 tokens/s on a base M4, which reads prompts at about 13 tokens/s. Runtimes downloaded before September 25 printed gibberish on M1 to M4; if that happened to you, download uzu/uzu again. The uzu/ folder holds a prebuilt runtime, so there is nothing to build:

hf download trymirai/Qwen3.8-27B-S-experimental --exclude "vllm/*" --local-dir qwen3.8-s
chmod +x qwen3.8-s/uzu/uzu
qwen3.8-s/uzu/uzu server --model qwen3.8-s

That starts an OpenAI-compatible endpoint on port 8000:

curl -s localhost:8000/v1/chat/completions -H 'Content-Type: application/json' -d '{"model":"qwen3.8-s","messages":[{"role":"user","content":"Capital of Australia? One word."}],"enable_thinking":false,"max_tokens":8}'

To chat in the terminal instead, leave out server. With the server running, python3 qwen3.8-s/speedcheck.py measures your Mac in under a minute; on an M5 Pro it prints 52 tokens/s decoding code (17 on prose), 440 tokens/s prefill, and 0.4 s to start answering a follow-up in a 7k-token chat. Follow-ups stay fast because uzu keeps the conversation cached between turns. More in uzu/README.md.

To build uzu yourself instead, install Rust with rustup, clone uzu's ryan/qtip-s-agent branch, and run cargo run --release -p cli -- server --model <model dir> inside it. If the build stops because the Metal toolchain is missing, run xcodebuild -downloadComponent MetalToolchain and try again.

Thinking is on by default. Turn it off per request with "enable_thinking": false, or server-wide by starting the server with UZU_SERVER_THINKING=0. Things to know:

• uzu's levels are low / medium / high / xhigh.
• With UZU_SERVER_THINKING=0, send enable_thinking together with reasoning_effort; reasoning_effort on its own returns 400.
• Streaming replies carry the thinking as reasoning_content deltas, then the answer as content, the way vLLM's reasoning parser does.

NVIDIA — vLLM

The vllm/ folder (9.3 GB download) holds a vLLM-ready copy of this checkpoint and the mirai_s plugin, which runs the compressed weights directly on the GPU: 8.2 GiB of GPU memory, so a 12 GB card runs it and a 16 GB card has room for 54k tokens of context with speculative decoding. On an RTX 3090 it answers one chat at 86 (prose) to 141 (code) tokens/s with speculative decoding, 135 tokens/s while reasoning, serves 620 tokens/s across 64 parallel requests, and has room for about 170k tokens of context. Install and serve commands, including one for drivers older than 580, are in vllm/README.md.

Known issues

• Both runtimes: a tight max_tokens with thinking on can spend the whole budget inside the trace and return an empty answer. Give reasoning requests at least 4000 tokens.