Qwen3.8-27B-FP8

FP8 quantization of Qwen3.8-27B, prepared and served by VIDRAFT.

Original weights, unmodified in behaviour. Quantized to FP8 for efficient high-throughput serving with the full 262,144-token context window intact.

Base model	Qwen/Qwen3.8-27B
Weights	FP8
KV cache	FP8
Context	262,144 tokens
Parameters	27B (64 layers, hidden 5120, 24 heads / 4 KV heads)
Languages	English, Korean, Chinese, Japanese, and more
License	Apache-2.0 (inherited)

---

Why serve this from VIDRAFT

We are the team that took #1 on the Google x Hugging Face "The Fast Gemma Challenge" leaderboard as of August 2, 2026, with a verified record of 510.58 tok/s.

What that number means:

• The challenge's own baseline - Google's official QAT W4A16 build, measured on the same harness and the same hardware - runs at 95.36 tok/s. Our entry reached 510.58 tok/s: a 5.35x speed-up.
• It came from serving optimization alone. No weight retraining, no distillation of the target model, no custom silicon. Same checkpoint, same GPU.
• The challenge re-measures every submission and marks it invalid when the result does not reproduce. Of 718 submissions, only 40 were verified. Ours is one of them.

That is the engineering we apply to this endpoint.

Leaderboard is public and continuously updated; rankings move as new entries are verified. The figures above are the verified record as of the date stated.

---

Serving characteristics

• Full 262K context. Not a truncated or sliding-window approximation - the model is served at its native maximum.
• Prompt caching. Repeated prefixes (system prompts, tool definitions, long documents) are billed at a reduced rate and returned faster. Cached token counts are reported in the standard usage object, so cost is auditable per request.
• Reasoning content is separated. The model's chain of thought is returned in a dedicated field, not mixed into the answer. Clients that ignore the field get a clean response; clients that want it can display it.
• OpenAI-compatible API, including tool calling and streaming.

---

Usage

from openai import OpenAI

client = OpenAI(base_url="https://openrouter.ai/api/v1", api_key="...")

resp = client.chat.completions.create(
    model="qwen/qwen3.8-27b",
    messages=[{"role": "user", "content": "Summarise this contract."}],
)
print(resp.choices[0].message.content)

---

About VIDRAFT

VIDRAFT is a Korean AI company building sovereign foundation models and inference systems. Selected public results:

• Fast Gemma Challenge - #1 verified record as of 2026-08-02 (510.58 tok/s)
• AETHER - a from-scratch open foundation model family, with training data, code, and logs released
• Darwin - a reasoning model family
• Research platforms for drug and materials discovery

Links: vidraft.net | Hugging Face

---

Notes

• This repository contains quantized weights and configuration only.
• Behaviour follows the base model. Refer to Qwen's model card for capabilities, evaluation results, and intended use.
• Apache-2.0, inherited from the base model.