Qwen 3.8 27B

ReasoningVisionCodeFunction CallingWeb Searchfp8private
🧠 Try in Intelligence →Try on Venice.ai ↗
Quick reference
Qwen 3.8 27B — TLDR
  • 🏢 Alibaba's Qwen team; dense 27B vision-language model, Apache 2.0 licensed.
  • 📏 Native 262,144-token context window, extensible further via YaRN scaling.
  • 👁️ Native image and video understanding built into the same model.
  • 🧠 Flexible thinking control: reasoning can be turned on or off.
  • 🔧 Focused on coding, professional work, research and agentic tasks.
  • ⚡ Served here in FP8; official FP8 weights published by Qwen.
  • 🆕 Successor to the 27B dense model of the Qwen 3.6 line.
  • 📚 Roughly 496K Hugging Face downloads at this catalog snapshot.
💰 Pricing
$0.450 / $3.20
per 1M · input / output
📏 Context
262K tokens
📅 On Venice since
Aug 17, 2026
1 day ago
Provider

Alibaba Group is a Chinese multinational technology company founded in 1999 and headquartered in Hangzhou, Zhejiang. Originally built around e-commerce and cloud computing, Alibaba has become one of the most prolific contributors to open-weight AI research,…

Read full profile →
62 models on Venice
23 video · 22 text · 7 image · 6 inpaint · 2 embedding · 2 tts
Since Jan 11, 2025

About this model

Qwen 3.8 27B is the compact dense member of Alibaba's Qwen 3.8 generation, released in 2026 under the Apache 2.0 license. The published model card describes it as a natively multimodal causal language model: a single set of weights handles text alongside image and video inputs, with a native context window of 262,144 tokens that can be extended further using YaRN-style rope scaling. Qwen also publishes an official FP8 checkpoint, which is the quantization served on this catalog, here with reasoning, vision, function calling and web search enabled.

Relative to its same-family predecessor Qwen 3.6 27B, the model keeps the same 27-billion-parameter dense footprint but is positioned by Qwen around stronger coding, professional and research work, and long-horizon agentic execution, together with flexible thinking control — reasoning can be enabled or disabled per request rather than requiring separate instruct and thinking checkpoints, as earlier Qwen releases such as Qwen 3 235B A22B Thinking 2507 did. Independent third-party benchmark results are not cited here.

Within the current lineup it sits below the flagship Qwen 3.8 2.4T mixture-of-experts model and the hosted Qwen 3.8 Max, while offering native vision that smaller dense builds like Qwen 3.5 9B did not carry in this catalog. That makes it a practical choice when you want multimodal input, tool use and a very long context at a size that fits a single accelerator.

This About section is AI-generated from public sources (Claude Opus 5), with no human editing. It may contain inaccuracies — verify critical details against the sources listed above.

Data sources: Venice API · HuggingFace · Wikipedia — enrichment updated 12h ago