🧬 Qwen4 architecture preview — $0.15 in / $0.47 out

Qwen3.8 Flash

Alibaba (Qwen)'s low-cost, multimodal MoE model. With just 6B active parameters, it beats Qwen3.7-Plus on coding, agentic and office benchmarks. An open-weight counterpart (Flash-Next) previewing the next-generation "Qwen4" architecture ships alongside it.

$0.15
Input / 1M tokens
$0.47
Output / 1M tokens
256K→1M
Context window
64K
Max output tokens
Last updated: | Sources: Qwen Official Blog · QwenCloud Model Page · Hugging Face | All Pricing · Calculator · 日本語版

Pricing (QwenCloud, international)

Pricing for the production Qwen3.8-Flash as stated on Qwen's official blog (qwen.ai/blog, 2026-08-26). All prices in USD per 1M tokens.

ModelInputOutputNotes
Qwen3.8-Flash (production API)$0.15$0.47Production SKU qwen3.8-flash. 1M context, official built-in tools
Qwen3.8-Flash-Next (open-weight)Weights free on Hugging Face / ModelScope (self-host)
Qwen3.8-Max$2.00$6.00Flagship (2.4T MoE). ~13× Flash price
💡 Pricing note: Qwen's official blog states "0.15 USD input / 0.47 USD output". However, the official X post (@Alibaba_Qwen) lists "$0.16 / $0.47", and mainland-China pricing is 1 CNY / 3 CNY (~$0.149 / $0.447, per Reuters). Always confirm final rates on the QwenCloud pricing page before contracting.
⚠️ Context differs by variant: the open-weight Flash-Next natively supports 262,144 (256K) tokens, extensible to 1M with YaRN. The production API Qwen3.8-Flash ships with 1M context by default (per the official blog). Self-hosting and API use differ in available context length.

Primary source: qwen.ai/blog?id=qwen3.8-flash-next (verified 2026-09-01)

Model Specifications

Release Date
Aug 26, 2026
Model ID (API)
qwen3.8-flash
Context Window
256K native → 1M (API: 1M default)
Max Output
64K tokens
Parameters
125B total / 6B active (+51B N-gram)
Input Modalities
Text + Image (+ video understanding)
Output Modality
Text
Architecture
GDN + QSA hybrid (sparse MoE)
Reasoning Effort
low / medium / xhigh (3 levels)
License
Open-weight (HF / ModelScope)
Provider
Alibaba (Qwen)

Sources: Qwen blog "Qwen3.8-Flash-Next: A New Architecture, Towards Ultimate Cost-Efficiency" · HuggingFace: Qwen/Qwen3.8-Flash-Next (verified 2026-09-01)

Performance — Beats Qwen3.7-Plus with 6B active

Figures below are self-reported by Qwen's official blog (2026-08-26) and may diverge from independent evaluations.

  • Coding: SWE-bench Pro 62.5 (Claude Opus 4.6 Max 53.4, Qwen3.7-Plus 55.8, DeepSeek-V4-Flash 56.0). DeepSWE 1.1 at 58.7 (Qwen3.7-Plus 16.5).
  • Office & agents: CoWorkBench 73.9 (Claude Opus 4.6 68.2), JobBench 55.7 (Qwen3.7-Plus 27.6).
  • Tool use: Toolathlon Verified (Pass@1) 73.5 (Qwen3.7-Plus 50.6, DeepSeek-V4-Flash 70.3).
  • Multimodal: AndroidWorld 84.5 (Claude Opus 4.6 62.0), LVBench 76.6 (Claude Opus 4.6 63.0), RealWorldQA 88.5 (Claude Opus 4.6 73.9).
  • General reasoning: GPQA Diamond 91.7 (Claude Opus 4.6 91.3), LiveCodeBench v6 91.9 (Claude Opus 4.6 88.8).
  • vs Qwen3.7-Plus: ~1/9 the training cost while outperforming it on major coding, office and agentic benchmarks.

Qwen positions the model as "a new architecture towards ultimate cost-efficiency." Via QSA (Qwen Sparse Attention), the Attention Kernel is up to 7.6× (Prefill) and 4.9× (Decode) faster at 1M tokens, and under a 90% Prefix Cache hit rate it reports 8.6× the Prefill throughput of Qwen3.7-Plus. Qwen released the weights early "so that the community can examine the architecture before the full Qwen4 model family is built on top of them." The production model now powers QwenWork's "Standard" mode.

Source: Qwen official blog (verified 2026-09-01)

Competitor Price Comparison

Comparison against low-cost Chinese and lightweight models. USD per 1M tokens (standard text, no cache).

ModelProviderInputOutputvs FlashNotes
Qwen3.8-FlashAlibaba$0.15$0.47Baseline. Multimodal, Qwen4 preview
GLM-5.3-FlashZhipu (Z.ai)$0.075$0.25GLM cheaperPromo (thru 9/9; list $0.15/$0.50)
GPT-5.6 LunaOpenAI$0.20$1.20Qwen cheaperQwen 25% cheaper input, 61% cheaper output
Gemini 3.1 Flash-LiteGoogle$0.25$1.50Qwen cheaperQwen 40% cheaper input, 69% cheaper output
DeepSeek V4 FlashDeepSeek$0.22–$0.44$0.66–$1.32ClosePeak/off-peak. Off-peak input higher than Qwen
Qwen3.8-MaxAlibaba$2.00$6.00Flash cheaperFlagship. Flash ~1/13 the price

* GLM-5.3-Flash is at 50%-off promo pricing (list $0.15/$0.50, thru 9/9). DeepSeek V4 Flash uses peak/off-peak pricing (cache-miss input $0.22 off-peak / $0.44 peak, output $0.66 off-peak / $1.32 peak). Qwen3.8-Flash is the international (USD) rate from the official blog.

Sources: official pricing pages / blogs from each provider (verified 2026-09-01). See full pricing comparison. GLM 5.3 Flash details →

Ideal Use Cases

Low cost × long context × multimodal workloads.

💻

Coding & Agents

SWE-bench Pro 62.5, DeepSWE 58.7. Works directly with Claude Code, Codex and Qoder CLI.

📄

Office & Document Workflows

CoWorkBench 73.9, JobBench 55.7. Finance, legal, medical and productivity tasks.

🔧

Tool Use & Function Calling

Toolathlon 73.5. OpenAI-compatible (Chat Completions / Responses) plus Anthropic-compatible protocols.

🖼️

Multimodal Understanding

Image and video input. AndroidWorld 84.5, LVBench 76.6 — strong on long video and GUI tasks.

🌐

Local Self-Hosting

Open weights on HF / ModelScope. Only 6B active enables memory-efficient inference.

🏗️

Early Qwen4 Evaluation

Preview the next-gen architecture (GDN + QSA) early — informing future Qwen4 adoption decisions.

Qwen 3.8 Family Selection Guide

ModelInputOutputWhen to Use
Qwen3.8-Max$2.00$6.00Hardest tasks, flagship quality
Qwen3.8-Flash ⬅$0.15$0.47Cost-sensitive, high-throughput, multimodal, coding
Qwen3.8-27BLocal inference, small dense model (license: see official model card)

💡 Strategy: Handle routine coding, agentic, document and multimodal work with Flash; escalate only the hardest tasks to Qwen3.8-Max.

Estimate Your Actual Costs

Compare token costs for Qwen3.8-Flash and all major models. Just enter your token counts.

🧮 Open Token Calculator →

📊 Full Pricing Comparison · 🇨🇳 China AI Overview · 🗓️ Release Timeline · 🔗 QwenCloud Model Page

⚠️ Disclaimer

  • Pricing verified against Qwen's official blog (qwen.ai/blog) on 2026-09-01. Subject to change without notice. Always check the QwenCloud pricing page before contracting.
  • Pricing differs between the official blog ($0.15/$0.47) and official X post ($0.16/$0.47); mainland China is priced in CNY (1/3). Rates may vary by region and contract terms.
  • Benchmark figures are self-reported by Qwen's official blog and may diverge from independent evaluations. Your mileage may vary.
  • This site is informational only. We are not affiliated with or endorsed by any provider. Qwen and Alibaba are trademarks of Alibaba Group.