LLMPrivate

Qwen 3.8 2.4T

Alibaba's open-weight, 2.4T parameter sparse MoE flagship — strong in coding, research, and long-context reasoning with 262K token context on Venice.

Maker
Alibaba Cloud
Modality
Text
Context
262.144K tokens
Open weights
Yes — under custom qwen3.8-max license

Overview

What is Qwen 3.8 2.4T

Qwen 3.8 2.4T is Alibaba's flagship open-weight language model, released in August 2026. Built on a sparse Mixture-of-Experts architecture with 2.4 trillion total parameters and 95 billion active per token, it excels in coding, research, and long-horizon agentic tasks with a 262,144-token context window and support for reasoning control and web search.

Running it privately on Venice

On Venice, Qwen 3.8 2.4T runs with full privacy — your prompts are never stored, profiled, or used for training. This uncensored, open-weights model is available without surveillance, giving developers and researchers sovereign control over sensitive workflows. You get the full power of Alibaba’s latest open flagship, now with zero retention and permissionless access.

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

Assessment

Strengths and limitations

Strengths
  • Exceptional performance in coding, research, and long-horizon agentic tasks, with strong results on benchmarks like PaperBench (93.0) and OSWorld (86.1).
  • Supports function calling, web search, and code-optimized generation, making it ideal for developer workflows and autonomous agents.
  • Open weights allow for self-hosting, fine-tuning, and full transparency — a rare trait among flagship models.
  • Large 262K context window with reasoning depth control via `reasoning_effort`, enabling complex, multi-step task execution.
Limitations
  • Text-only input: does not support vision or multimodal inputs despite the multimodal Qwen3.8-Max variant.
  • Always runs in reasoning mode: cannot disable chain-of-thought, which may increase latency and cost for simple queries.
  • Lower GPQA Diamond score (92.6) compared to GPT-5.6 Sol (94.1), indicating a slight gap in factual reasoning.

Capabilities

What it supports

  • Tool use / function calling
  • Reasoning
  • Web search
  • Code-optimized generation
  • Vision input
  • Video input
  • Non-thinking mode
  • Fine-tuning support

Specifications

Datasheet

Maker
Alibaba Cloud
Open weights
Yes — under custom qwen3.8-max license
License
Custom: qwen3.8-max
Modes
Reasoning modes: low, medium, xhigh; always reasons
Context window
262.144K tokens
Prompt length
262.144K tokens
Released
August 2026
Architecture
Sparse Mixture-of-Experts (MoE)
Parameters
2.4T total, ~95B activated per step
Max output
65.536K tokens
Capabilities
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-3-8-2-4t-a95b",
    "messages": [{ "role": "user", "content": "Explain quantum tunneling simply." }]
  }'

Pricing

What it costs on Venice

Billed per token on Venice: $2.50 per 1M input tokens and $7.50 per 1M output tokens.

Input / 1M tokens
$2.50
Per 1M tokens
Output / 1M tokens
$7.50
Per 1M tokens
Cached input / 1M
$0.31
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 2.4TOpen flagship coding & research262K tokensYes$2.50 in · $7.50 out / 1M
Claude Fable 5Proprietary reasoning1M tokensNo$12 in · $60 out / 1M
DeepSeek V4 Flash 0731Speed & cost1M tokensYes$0.17 in · $0.35 out / 1M
GLM 5.1Balanced open model200K tokensYes$1.10 in · $4.15 out / 1M

Choose Qwen 3.8 2.4T when you need the highest reasoning and coding performance among open, private models — ideal for developers and researchers who demand transparency and control.

Use cases

What it is good for

  1. 01Long-form code generation and debugging with access to external tools and documentation.
  2. 02Research summarization and synthesis across large document sets within the 262K context window.
  3. 03Autonomous agent workflows requiring reliable multi-step planning and environment interaction.
  4. 04High-stakes reasoning tasks where transparency and auditability are required via open weights.

Prompting

Getting better results

Use `reasoning_effort=low` for faster, cheaper responses on simple queries.

Set `preserve_thinking=true` to retain reasoning context across conversation turns.

Include explicit tool calls in your prompt if the model doesn’t auto-invoke them.

Structure complex tasks as step-by-step instructions to leverage its agentic strengths.

Use web search mode for up-to-date factual queries outside its training data.

Keep prompts under 262K tokens to maximize output space for long-form responses.

Version history

Qwen3.5
2025

Architectural foundation

Qwen3.6
2026-06

Stability and utility improvements

Qwen3.7 Max
2026-07

Predecessor flagship

Qwen3.8 2.4T
2026-08

Current — open flagship with 2.4T MoE

FAQ

Frequently asked questions

Qwen 3.8 2.4T is Alibaba's flagship open-weight language model, released in August 2026. It features a sparse Mixture-of-Experts architecture with 2.4 trillion parameters and excels in coding, research, and long-context reasoning tasks with a 262K-token context window.

The model weights are open and available under a custom license called 'qwen3.8-max', allowing self-hosting and inspection. However, it is not free to run — on Venice, it costs $2.50 per million input tokens and $7.50 per million output tokens.

The model supports a context window of 262,144 tokens, with up to 65,536 tokens for output. This allows for extremely long conversations and document processing.

No. The Qwen 3.8 2.4T model available on Venice is text-only. While the hosted Qwen3.8-Max variant supports multimodal inputs, this open checkpoint does not.

Yes. The model supports function calling and tool use, making it suitable for agentic workflows, code execution, and integration with external APIs.

Qwen 3.8 2.4T is open and private, while Claude Fable 5 is closed and more expensive. Fable 5 leads in SWE-bench and HLE, but Qwen offers superior cost efficiency and sovereignty for developers who need open models.

The model is open weights and uncensored in the sense that it does not enforce policy-based filtering by default. However, deployment platforms may apply their own content policies.

Run Qwen 3.8 2.4T privately

No prompt logging. No data used for training.