LLMPrivate

Qwen 3.8 27B

A compact, open-weight, multimodal 27B LLM excelling in coding, agentic workflows, and vision tasks — runs locally, scales to 1M context, and operates privately on Venice.

Maker
Alibaba
Modality
Text + vision
Context
262.144K tokens
Open weights
Yes — Apache 2.0

Overview

What is Qwen 3.8 27B

Qwen 3.8 27B is a 27-billion-parameter open-weight multimodal language model from Alibaba, released in August 2026 under Apache 2.0. It natively supports text, vision, and function calling, with a 262K-token context window extendable to 1M, optimized for coding, agentic tasks, and local deployment.

Running it privately on Venice

On Venice, Qwen 3.8 27B runs with zero retention — your prompts are never stored, profiled, or reused. This ensures full sovereignty over sensitive code, research, or business workflows. The model’s open weights and uncensored reasoning enable permissionless deployment in regulated or high-compliance environments.

Private (zero retention)No prompt trainingTEE · hardware enclaveEnd-to-end encrypted

Assessment

Strengths and limitations

Strengths
  • Excellent coding and software engineering performance — leads on SWE-bench Pro and QwenSWEBench.
  • Strong multimodal agent: handles image and video input natively, excels in OSWorld and AndroidWorld tasks.
  • Open weights under Apache 2.0: free for commercial use, self-hostable, and modifiable.
  • 262K native context, extendable to 1M tokens via YaRN — ideal for long-horizon agentic workflows.
  • Supports vision, function calling, reasoning, and web search — full-stack agentic capabilities.
Limitations
  • Default reasoning mode (xhigh) can lead to overthinking and slow responses on consumer hardware.
  • Local 1M context requires YaRN configuration: not automatic; hosted versions support it by default.
  • Not end-to-end encrypted or running in a TEE on Venice — privacy is zero retention but not hardware-isolated.

Capabilities

What it supports

  • Tool use / function calling
  • Vision (image input)
  • Reasoning
  • Web search
  • Code-optimized
  • Structured output (JSON schema)
  • Video input
  • Multiple image inputs
  • End-to-end encryption
  • Runs in TEE

Specifications

Datasheet

Maker
Alibaba
Open weights
Yes — Apache 2.0
License
Apache 2.0
Modes
Thinking (xhigh/medium/low), Instruct
Context window
262.144K tokens
Prompt length
Up to 262.144K tokens
Input images
Accepted: JPEG, PNG, WebP, GIF; supports video input up to hour-scale
Released
August 14, 2026
Architecture
Dense causal LLM with vision encoder, Gated DeltaNet and Gated Attention blocks
Parameters
27B
Max output
65.536K tokens
Capabilities
Vision, Function calling, Reasoning, Web search, Code-optimized
Privacy on Venice
Private — zero retention
Available on Venice since
Aug 2026

API

Call it from your code

Venice exposes an OpenAI-compatible API. Point your base URL at Venice and pass the model id.

curl https://api.venice.ai/api/v1/chat/completions \
  -H "Authorization: Bearer $VENICE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "qwen-3-8-27b",
    "messages": [{ "role": "user", "content": "Explain quantum tunneling simply." }]
  }'

Pricing

What it costs on Venice

Billed per token on Venice: $0.45 per 1M input tokens and $3.20 per 1M output tokens.

Input / 1M tokens
$0.45
Per 1M tokens
Output / 1M tokens
$3.20
Per 1M tokens

New Venice accounts include a free daily allowance and 500 welcome credits — no credit card required.

Alternatives

How it compares

ModelBest forContextOpen weightsPrice (Venice)
Qwen 3.8 27BCoding & agentic tasks262K tokensYes$0.45 in · $3.20 out / 1M
DeepSeek V4 Flash 0731Speed & cost efficiency1M tokensYes$0.17 in · $0.35 out / 1M
Google Gemma 4 31B InstructEfficient local inference256K tokensYes$0.12 in · $0.36 out / 1M
GLM 5.1Balanced multimodal200K tokensYes$1.10 in · $4.15 out / 1M

Choose Qwen 3.8 27B when you need a powerful, open, and vision-capable model for coding, agentic workflows, and local sovereignty — especially if you value Apache 2.0 licensing and multimodal reasoning.

Use cases

What it is good for

  1. 01Local coding agents that plan and execute software tasks autonomously.
  2. 02Multimodal research assistants analyzing documents, screenshots, and videos.
  3. 03Private business automation with vision and web tools, no data leakage.
  4. 04High-compliance environments requiring open, auditable, and self-hostable models.
  5. 05Long-context agentic workflows like legal analysis or technical documentation.

Prompting

Getting better results

Use reasoning_effort=medium or low for faster responses on local hardware.

Enable preserve_thinking to maintain context across conversational turns.

Provide images and video as direct input for multimodal reasoning tasks.

Use temperature 0.7 and top_p 0.8 for instruct mode; 1.0 and 0.95 for thinking mode.

Structure complex tasks with explicit step-by-step directives to guide agent execution.

Leverage web search capability when up-to-date information is required.

FAQ

Frequently asked questions

Qwen 3.8 27B is a 27-billion-parameter open-weight multimodal language model from Alibaba, released in August 2026. It supports text, vision, and function calling, with a 262K-token context window extendable to 1M, optimized for coding, agentic tasks, and local deployment.

Yes, it is open weights under the Apache 2.0 license, meaning it is free for commercial use, modifiable, and self-hostable. The model weights are available on Hugging Face.

On Venice, pricing is $0.45 per 1M input tokens and $3.20 per 1M output tokens. There are no subscription fees — you pay only for what you use.

Yes, it natively supports image and video input, including hour-scale videos, making it suitable for multimodal reasoning and agent tasks.

Yes, it supports function calling and tool use, enabling it to execute code, search the web, and interact with external systems as part of agentic workflows.

It has a native context window of 262,144 tokens and can be extended to 1 million tokens using YaRN scaling, ideal for long-horizon tasks.

Qwen 3.8 27B is stronger in coding, vision, and agentic tasks, while DeepSeek V4 Flash is faster and cheaper. Choose Qwen for capability, DeepSeek for efficiency.

Run Qwen 3.8 27B privately

No prompt logging. No data used for training.