MODEL RECORD · REPORTED + OBSERVED

Qwen3.8-27B

An immutable model record joined to reviewed Q4_K_M receipts. Performance is observed for one RTX 4090 configuration; quality remains unmeasured.

reportedReviewed Aug 18, 2026

What is established

Parameters
27.78B
Architecture
Qwen3_5ForConditionalGeneration
Transformer blocks
64
Full-attention blocks
16
Native context
262,144 tokens
FP16 KV cache
64 KiB / token
Repository revision
1d4bf0f2ff60
License
Apache-2.0
NEXT EVIDENCE

What still needs to be measured

Other quantizations

Repeat the fit boundary across Q5, Q6 and Q8 using pinned artifacts.

Other hardware

Desktop 4090, 5090, Macs and CPU-only systems need separate receipts.

Quality loss

A pre-registered evaluation across identical prompts and settings.

PROVENANCE

Primary sources

  1. Qwen/Qwen3.8-27B registry recordHugging Face / Qwen · retrieved Aug 18, 2026
    Open source ↗
  2. Qwen3.8-27B immutable configQwen · retrieved Aug 18, 2026
    Open source ↗