AI
AI Updates
Back to all posts
·
qwenmodel-releasesdistillationinference

Qwen 3.8: The 27B That Doesn't Exist Yet (and What Does)

Breaking down the Qwen 3.8 landscape — the official 2.4T Max, the missing 27B, and community distillation variants.

A
AI Updates

The Qwen 3.8 Family

Alibaba announced the Qwen 3.8 generation on August 3, 2026. Here’s what actually exists as of August 12:

Official Models

Model Parameters Architecture Status
Qwen 3.8-Max 2.4T (95B active MoE) Sparse MoE ✅ Released on HuggingFace
Qwen 3.8-Max FP8 2.4T FP8 quantized Sparse MoE ✅ Released
Qwen 3.8-27B ~27B dense Dense ❌ Not yet released

The 27B Situation

The Qwen 3.8-27B-Instruct has been announced but has not been released on HuggingFace. The official Qwen/Qwen3.8-27B-Instruct endpoint returns nothing.

This follows the same pattern as previous generations:

  • Qwen 3.6-27B: 2,241 likes on HF
  • Qwen 3.5-27B: 1,029 likes on HF

Both were released alongside their Max counterparts with a short delay.

Community Distillation Variants

While the official 27B isn’t here yet, the community has already started distilling the Qwen 3.8-Max:

Model Author Size Likes Notes
Qwen3.8-27B-FP8 huginnfork 27B FP8 22 FP8 quantized variant
Qwen3.8-27B-NVFP4A16 huginnfork 27B NVFP4 10 NVIDIA FP4 format
Qwen3.8_4B_Distilled Ma7ee7 4B 20 Distilled from Qwen3.8-Max preview
Qwen3.8-27B-GGUF barozp 27B 1 GGUF conversion
Qwen3.8-27B-Opus-Distill-GGUF barozp 27B 2 Distilled from Claude Opus?

Important: These are community projects, not official Qwen releases. The huginnfork variants appear to be quantization experiments rather than true knowledge distillation. The Ma7ee7 4B model is a genuine distillation attempt from Qwen 3.8-Max preview outputs.

What’s the Difference?

The 27B is NOT a distilled version of the 2.4T Max. It’s a separate dense model line:

  • Qwen 3.8-Max (2.4T): Sparse MoE — 2.4T total parameters, ~95B active per token. API-only deployment.
  • Qwen 3.8-27B (expected): Dense — ~27B total parameters. Designed for self-hosting on single GPUs (17GB+ VRAM at 4-bit).

Performance Expectations

Based on the Qwen 3.6-27B → 3.5-27B upgrade pattern, the 3.8-27B is expected to bring:

  • Improved agentic coding capabilities
  • Better context handling
  • Deeper reasoning chains
  • Compatibility with existing Qwen tooling

VRAM Requirements

Quantization RTX 3060 (12GB) RTX 3090 (24GB)
Q4_K_M ❌ Too large ✅ Comfortable
Q5_K_M ❌ Too large ✅ Good margin
Q6_K ❌ Too large ✅ Tight fit
Q8_0 ❌ Impossible ⚠️ Bare minimum

What to Watch

  1. Official 27B release — Expected within days based on the Qwen 3.6 pattern
  2. Community GGUF conversions — Already appearing from unsloth and others
  3. Benchmark comparisons — LMSYS Arena entries pending
  4. Ollama integration — Will follow HF release by ~48 hours

Sources