Qwen3.8 Flash

Fast hybrid thinking model with strong agentic coding and 1M context

Try now

read docs

About model

Qwen3.8 Flash is the production version of Qwen3.8-Flash-Next, the model Alibaba built as an early preview of the Qwen4 architecture. It pairs a 125B parameter main model activating just 6B per token with an additional 51B parameters of n-gram embedding memory, expanding capacity with almost no extra computation per token. The architecture combines Gated DeltaNet, which compresses history efficiently, with Qwen Sparse Attention, which retrieves important context at micro-block granularity, keeping the 1M-token context window fast and affordable. Qwen reports training cost at roughly one-ninth that of Qwen3.7-Plus while delivering stronger coding and office-task results, and the model carries upgraded vision-language understanding with reasoning effort adjustable per request. Available on Together AI.

‍

Activated Parameters

125B main model plus 51B of n-gram embedding memory

Context Window

Hybrid sparse attention designed to keep long context fast and affordable

Training Cost vs Qwen3.7-Plus

1/9

Qwen-reported, with stronger coding and office-task results

Model key capabilities

Agentic Coding & Tool Use: Strong results across real-repository engineering, multilingual software work, and multi-step tool workflows
Vision-Language Understanding: Upgraded image understanding across charts, documents, interfaces, and real-world scenes
Ultimate Cost Efficiency: Hybrid Gated DeltaNet and sparse attention with n-gram embedding memory, built to produce more intelligence with less compute
Available on Together AI: Serverless access with a 1M-token context window through the Together AI API

Performance benchmarks

Model	FrontierMath Tier 4	GPQA Diamond	HLE	SciCode	GDPval-AA	Terminal-Bench 2.1	Agent Arena	FrontierCode	DeepSWE
Qwen3.8 Flash		91.7	35.9
Related open-source models
Competitor closed-source models
Claude Fable 5	87.8%	92.6%	53%	60%	62%	85%		53.5%	70%
Claude Opus 5	73.2%	93.2%	53%	56%	68%	89%	+16.4pp	53.4%	74%
GPT-5.6 Sol	82.9%	94.1%	47%	56%	61%	88%		47.5%	73%
Grok 4.5	24.4%	93.1%	40%	54%	51%	82%	+4.2pp	42.4%	54%
GPT-5.6 Luna	61.0%	91.1%	37%	53%	54%	81%		39.8%	67%

API usage

cURL
Python
Typescript

Endpoint:

Qwen/Qwen3.8-Flash

curl -X POST "https://api.together.xyz/v1/chat/completions" \
  -H "Authorization: Bearer $TOGETHER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "Qwen/Qwen3.8-Flash",
    "messages": [
      {
        "role": "user",
        "content": "What are some fun things to do in New York?"
      }
    ]
}'

from together import Together

client = Together()

response = client.chat.completions.create(
  model="Qwen/Qwen3.8-Flash",
  messages=[
    {
      "role": "user",
      "content": "What are some fun things to do in New York?"
    }
  ]
)
print(response.choices[0].message.content)

import Together from 'together-ai';
const together = new Together();

const completion = await together.chat.completions.create({
  model: 'Qwen/Qwen3.8-Flash',
  messages: [
    {
      role: 'user',
      content: 'What are some fun things to do in New York?'
     }
  ],
});

console.log(completion.choices[0].message.content);

Model card
Architecture Overview:
• 125B parameter main model with 6B activated per token, plus 51B n-gram embedding parameters providing lookup-based local-pattern memory at negligible per-token compute
• Hybrid attention: three of every four layers use Gated DeltaNet to compress history into a fixed-size state, while the fourth uses Qwen Sparse Attention (QSA), a lightweight indexer that aggregates the sequence into micro-blocks and selects the most relevant regions
• Gated Residual widens the residual stream into four dynamically gated branches, strengthening cross-layer information flow and training stability, with FP8 residual-state storage
• Ultra-sparse MoE with a large expert pool, few routed experts per token, and one shared expert; Multi-Token Prediction module with QSA in its attention layers
• 262,144-token native context, served in production with a 1M-token context window

Training Methodology:
• Released as an early preview of the Qwen4 architecture, the same role Qwen3-Next played ahead of the Qwen3.5 through Qwen3.8 series
• Trained with the Muon optimizer for two-dimensional linear weights and AdamW for embeddings, the MoE router, and low-rank parameters, with the scaling law refitted for the new architecture
• Qwen reports training cost at roughly one-ninth that of Qwen3.7-Plus

Performance Characteristics:
• Agentic coding (DeepSWE 1.1: 58.7, versus 42.2 for Qwen3.8-27B; SWE-bench Pro: 62.5; SWE-bench Multilingual: 81.0; NL2Repo repository generation: 48.1)
• Long-horizon professional work (CoWorkBench, Qwen in-house: 73.9; JobBench professional tasks: 55.7, versus 33.4 for Qwen3.8-27B; Agents' Last Exam: 24.3 pass@1)
• Tool use (Toolathlon Verified: 73.5); instruction following (IFBench: 81.3)
• Reasoning and knowledge (GPQA Diamond: 91.7; Humanity's Last Exam: 35.9; LiveCodeBench v6: 91.9)
• Vision (AndroidWorld mobile use: 84.5; Vision2Web visual web development: 64.0; RealWorldQA: 88.5; MathVision: 90.6, or 95.7 with code interpreter; CharXiv RQ chart reasoning: 84.6; LVBench long video understanding: 76.6)
• Qwen reports the QSA attention kernel reaches up to 7.6x prefill and 4.9x decode speedups at 1M tokens, and 8.6x the prefill throughput of Qwen3.7-Plus at 1M context under high cache reuse
‍
Prompting
Together AI API Access:
• Access Qwen3.8 Flash via Together AI APIs using the endpoint Qwen/Qwen3.8-Flash
• Authenticate using your Together AI API key in request headers
• Reasoning effort is adjustable per request across xhigh, medium, and low levels
• Supports streaming responses across a 1M-token context window
• Available on Together AI with serverless access
‍
Applications & use cases
High-Volume Agent Workloads:
• Serve tool-driven agent traffic at a 6B-active-parameter cost profile
• Dial reasoning effort down for routine calls and up to xhigh for hard steps
• Hold long agent trajectories in the 1M-token window across sessions

Coding & Coworking Assistants:
• Run repository-scale engineering and multilingual software work through the Together endpoint
• Power office and productivity agents spanning documents, finance, legal, and analysis tasks
• Build coding assistants where per-request cost matters as much as capability

Long-Context & Visual Analysis:
• Cross-reference large codebases and document sets in a single request
• Reason over charts, documents, interfaces, and real-world images alongside text
• Combine visual math and chart analysis with long-context working sets
‍

Related models

Model specifications

Model data

Model provider
Qwen
Type
Chat
Reasoning
Vision
Deployment
Serverless
Endpoint
Qwen/Qwen3.8-Flash
Parameters
125B
Activated parameters
6B
Context length
1M
Input price
$0.15 / 1M tokens
Output price
$0.47 / 1M tokens
Input modalities
Text
Image
Output modalities
Text

Released
August 25, 2026
Category
Chat

Run in Playground

Quickstart docs

Deploy model

Qwen3.8 Flash

About model

API usage

Model card

Prompting

Applications & use cases