Robots Atlas>ROBOTS ATLAS
Qwen3.8-Flash-Next
AI Modelsโ€บQwen3

Qwen3.8-Flash-Next

3.8-Flash-Nextย ยทย Family: Qwen3
Qwen (Alibaba) hybrid MoE multimodal model: 125B params, 6B active, hybrid Gated DeltaNet + Qwen Sparse Attention, 262K context (up to 1M).
โœ“ Activeโœ“ Public accessโš– Open weightsMultimodalLLMReasoning model๐Ÿ“ Qwen3
Context window
262K
tokens
Parameters
125B (6B active)
parameters
Release date
1 August 2026
๐ŸขQwenProducer
Access:APIDownloadHostedDeployment:๐Ÿ’ป Localโ˜ Cloud

Overview

Qwen3.8-Flash-Next is a Qwen-family multimodal language model developed by the Qwen team (Alibaba), released in August 2026 on Hugging Face. It uses a hybrid Mixture-of-Experts architecture: 125B total parameters and 6B active per token, 48 layers and 512 MoE experts (10 routed + 1 shared).

The model combines two attention mechanisms at the micro-block level โ€” Gated DeltaNet and Qwen Sparse Attention (QSA) โ€” in a "Hybrid Attention with QSA" scheme. It introduces "Gated Residual" for modulated information flow across widened residual streams and an N-gram embedding approach (51B parameters) to scale parameters on memory-constrained accelerators; it also uses an MTP module (4B). The training recipe combines the Muon and AdamW optimizers. It supports a thinking mode with configurable reasoning effort.

The native context window is 262,144 tokens, extensible to 1,000,000 tokens via RoPE scaling. The model accepts text, image and video. Weights are released under the qwen-community-1.0 license.

Classification
MultimodalLLMReasoning model
Family: Qwen3
Access & deployment
APIDownloadHosted
LocalCloud
Weights: Open weights
Key parameters
๐Ÿ“ Context: 262K
๐Ÿงฉ Parameters: 125B (6B active)
โœ“ Toolsย ยทย โœ“ Fine-tuning
๐Ÿ“ฅ Input: text, image, video

Technical specification

Context window
262K
tokens
Parameters
125B (6B active)
parameters
License
qwen-community-1.0
Hardware requirements
Open weights on Hugging Face; the N-gram embedding approach targets memory-constrained accelerators. The 125B MoE model requires multi-GPU class resources.
Features:โœ“ Tool useโœ“ Fine-tuning
Modalities
โฌ‡ Input
textimagevideo
โฌ† Output
textcode

Capabilities and applications

Native model capabilities
Reasoning
The model's ability to reason logically and solve complex problems.
Category: reasoning
Multi-step reasoning
Carrying out multi-step chains of reasoning across long, complex tasks.
Category: reasoning
Mathematical reasoning
The model's ability to solve mathematical tasks requiring multi-step reasoning โ€” equations, proofs, combinatorics, geometry, calculus and competition-level problems.
Category: reasoning
Coding
Generating, analysing and modifying code in many programming languages. Covers writing functions, debugging, refactoring, code review, and creating tests. Measured by benchmarks such as HumanEval and SWE-bench.
Category: coding
Agentic coding
Multi-hour, multi-step programming tasks performed autonomously by the model: cloning a repository, running tests, iterating on fixes, integrating with CLI tools. Characteristic of Codex variants (GPT-5.1-Codex-Mini, Codex-Max).
Category: coding
Agentic capability
The model's ability to autonomously plan and execute multi-step tasks by sequentially using tools, maintaining context, and adapting to intermediate results.
Category: planning
Tool use
The model's ability to call external functions, APIs and tools during a conversation: calculator, search engine, code editor, database. The model decides when and how to use a tool and interprets its result.
Category: planning
MCP support
Native support for the Model Context Protocol - the model can integrate with external MCP servers, invoke their tools, and use their data sources without a dedicated wrapper.
Category: other
Long context
Support for large context windows โ€” tens to hundreds of thousands (or millions) of input tokens. Enables analysis of entire codebases, long documents, and many parallel conversations without losing earlier information. GPT-5.1 supports 400,000 tokens.
Category: language
Multilingual
Competence in many natural languages (from a few to over a hundred): understanding, generation, translation, and code-switching within a single conversation. Frontier models support a wide range of languages with comparable quality.
Category: language
Structured output
Producing data in structured formats such as JSON.
Category: structured_generation
Extended thinking mode
A reasoning-model variant with a larger inference budget: more thinking cycles, higher answer precision at the cost of response time. Choice between 'standard' and 'extended' thinking is left to the user (e.g. the selector in GPT-5.2 Pro).
Category: reasoning
Adaptive reasoning effort
The model decides how much 'thinking' to allocate to a given query: simple questions are answered quickly, complex problems receive more inference cycles. A GPT-5.1 feature (both Instant and Thinking) that shortens time on easy tasks and extends it for hard ones.
Category: reasoning
Language modeling
Ability to predict subsequent tokens and generate coherent natural-language text based on the preceding context.
Category: language

Benchmark results

5 benchmarks
DeepSWE 1.1
58.7%
๐Ÿ“„ Karta modelu Hugging Face (Qwen/Qwen3.8-Flash-Next)
SWE-bench Pro
62.5%
๐Ÿ“„ Karta modelu Hugging Face (Qwen/Qwen3.8-Flash-Next)
SWE-bench Multilingual
81.0%
๐Ÿ“„ Karta modelu Hugging Face (Qwen/Qwen3.8-Flash-Next)
GPQA Diamond
91.7%
๐Ÿ“„ Karta modelu Hugging Face (Qwen/Qwen3.8-Flash-Next)
ClawEval-MM
64.4%
๐Ÿ“„ Karta modelu Hugging Face (Qwen/Qwen3.8-Flash-Next)
Pass@3.

Technical architecture

Sources and related pages