Qwen3.8-Flash Review: $0.15/M Input, 67% Cheaper Than DeepSeek | Lieke Tech
- # Qwen3.8-Flash: Full Review
Price cut just one day after launch — $0.15/M input tokens, 89% lower training cost, coding benchmarks match DeepSeek V4 Pro
✅ Input $0.15/M tokens
✅ Output $0.47/M tokens
✅ Cache hit $0.016/M
Claim $200 Free Credit →
ECS from $4.50/mo
📌 Key Takeaways
Launch then price cut: Qwen3.8-Flash launched on August 26 at $0.16/$0.47 per million tokens; Alibaba cut the China-side price on August 27 to ¥0.8 (input) and ¥2.7 (output), with cache hits at ¥0.1/M
Architecture breakthrough: 125B-parameter MoE model activating only 6B per token, plus a 51B N-gram embedding layer — training cost is ~1/9 of Qwen3.7-Plus (89% reduction)
Competitive performance: Scores 62.5 on SWE-bench Pro (agentic coding), leads in 8 of 14 benchmarks, matching DeepSeek-V4-Flash and Claude Opus 4.6 on real-world tasks
Mega context: Natively supports 262K tokens, extends to 1M with YaRN. Multimodal: text, image, and video input
Cache economics: Cache hit at $0.016/M tokens — ideal for agents, RAG, and long conversations
💰 Latest Pricing (Effective August 27, 2026)
Pricing via Alibaba Cloud Model Studio (international) and Bailian platform (China), per million tokens.
International Pricing (USD)
| Billing Item | Price (per 1M tokens) |
|---|---|
| Input | $0.15 |
| Output | $0.47 |
| Cache hit | $0.016 |
China Pricing (RMB, after Aug 27 cut)
| Billing Item | Launch Price (Aug 26) | Current Price (Aug 27+) | Change |
|---|---|---|---|
| Input | ¥1.0 | ¥0.8 | ↓20% |
| Output | ¥3.0 | ¥2.7 | ↓10% |
| Cache hit | — | ¥0.1 | — |
💡 Real-world cost example: An AI customer service app handling 1M requests/day (avg 2K input + 500 output tokens):
Qwen3.8-Flash: ~$415/day total
DeepSeek V4 Flash (peak): ~$1,260/day
Savings: ~$300,000+ per year
⚔️ Competitive Pricing Comparison
August 2026 pricing for major LLM APIs, USD per million tokens.
| Model | Input | Output | Active Params | Context |
|---|---|---|---|---|
| Qwen3.8-Flash NEW | $0.15 | $0.47 | 6B (MoE) | 1M |
| Qwen3.5-Flash (≤128K) | $0.11 | $0.67 | — | 1M |
| Qwen3.7-Flash | $0.034 | $0.132 | — | 256K |
| Qwen3.8-Max | $2.00 | $6.00 | — | 1M |
| DeepSeek V4 Flash (peak) | $0.42 | $1.26 | 13B | 128K |
| DeepSeek V4 Flash (off-peak 50%) | $0.21 | $0.63 | 13B | 128K |
| DeepSeek V4 Pro (peak) | $1.26 | $3.78 | — | 128K |
| GLM-5.3-Flash | ~$0.15+ | ~$0.47+ | 18B (MoE) | — |
Key findings:
Qwen3.8-Flash output is 63% cheaper than DeepSeek V4 Flash peak ($0.47 vs $1.26)
Even against DeepSeek's off-peak pricing, Qwen3.8-Flash remains 25% cheaper on output
API price is 1/10 of GLM-5.3, and up to 1/20 during promotional periods
Qwen3.7-Flash is cheaper but lacks the 1M context and multimodal capabilities of 3.8
🏗️ Architecture Deep Dive
Qwen3.8-Flash (codenamed Qwen3.8-Flash-Next during development) is an early preview of the Qwen4 architecture, with four core innovations:
1. Mixture-of-Experts (MoE)
125B main parameters with only 6B activated per token (10 routed experts + 1 shared expert out of 512)
Additional 51B N-gram embedding layer stored in system RAM (not GPU), acting as a massive "phrase dictionary"
4B multi-token prediction parameters bring total footprint to ~180B
2. QSA Sparse Attention + GDN Hybrid
Qwen Sparse Attention (QSA) works alongside GDN: GDN compresses historical context while QSA selects key information. In high-cache-hit 1M-token scenarios, this delivers 7.6x faster prefill and 4.9x faster decode.
3. Gated Residual Mechanism
Splits the traditional single residual pathway into 4 parallel branches, dynamically gating information flow for better cross-layer communication and training stability.
4. Muon Optimizer
Uses a refined Muon + AdamW hybrid strategy with refitted scaling laws, eliminating batch warm-up and significantly improving convergence efficiency and training throughput.
Result: Training costs are approximately 1/9 of Qwen3.7-Plus (397B params, 17B activated) — an 89% reduction — while delivering superior coding and office task performance.
📊 Benchmark Performance
Per Qwen's official technical report, Qwen3.8-Flash excels across multiple evaluation suites:
| Benchmark | Domain | Performance |
|---|---|---|
| SWE-bench Pro | Agentic coding | 62.5, leading peers |
| CoWorkBench | Long-horizon office tasks | Beats DeepSeek V4 Flash |
| Toolathlon Verified | Real-world tool use | Matches Claude Opus 4.6 |
| MathVision | Visual math reasoning | Strong multimodal |
| AndroidWorld | Mobile agent tasks | Embodied intelligence |
| ERQA | Embodied reasoning | Multimodal understanding |
Across 14 evaluations, the base model (6B activated) achieved the best results in 8. The fine-tuned version shows even stronger performance in coding, agents, and multimodal tasks.
🎯 Best Use Cases
Qwen3.8-Flash is ideal for:
AI coding assistants: 62.5 on SWE-bench Pro, repository-level understanding with 1M context
AI agents / tool use: Native parallel tool calls, extremely low cost for long-horizon tasks
Customer service / chatbots: $0.016/M cache hit makes multi-turn conversations nearly free
Document analysis / RAG: 1M tokens processes hundreds of pages, cache speeds up 8x
Multimodal apps: Text + image + video input, visual math and chart analysis
High-volume batch processing: 5,000 RPM / 5M TPM rate limits for enterprise scale
🚀 How to Get Started
Option 1: Alibaba Cloud Model Studio (International)
Access Qwen3.8-Flash via Model Studio with OpenAI-compatible API. New users get free credit. The model serves on QwenCloud with 1M context by default.
Claim $200 Free Credit →
Option 2: Bailian Platform (China)
Available on Alibaba Cloud's Bailian platform with new user free token quota. First to receive the latest Qwen releases.
Bailian Console →
Option 3: Open Weights (Self-hosted)
The open-weight version, Qwen3.8-Flash-Next, is available on Hugging Face and ModelScope for download, fine-tuning, and local deployment. The production version with built-in tools and 1M default context is served via QwenCloud API.
New user offers:
Alibaba Cloud international: up to $200 free trial credit
Bailian platform: free token quota for new users (90 days)
ECS cloud servers: from ~$4.50/month (2 vCPU, 2GB RAM)
📰 September 19, 2026 update: Qwen3.8-27B generates working web apps from a single prompt
Qwen3.8-27B keeps trending in developer circles this week. In a hands-on test published by QbitAI on September 19, a single prompt led Qwen3.8-27B to produce a 52KB offline data-analysis tool in about 5 minutes — automatically cleaning tables, aggregating revenue stats and rendering bar charts — and even a realistic 12306 ticket-booking page (a convincing front-end shell that, by default, does not connect to live data or payment backends). Engineer Alok paired Qwen3.8-27B with Cerebras to build an offline "AI computer desktop" that regenerates historical versions of Google, YouTube and other sites on the fly at roughly 1,950 tokens/s (individual benchmark). On September 18, the Qwen team also launched Qwen3.8-Omni-Flash, a next-generation omnimodal model with text/image/audio/video inputs and a 1M-token context, cutting per-hour audio-input API pricing by more than 98% versus the previous generation. From local 27B deployment to cloud APIs, the Qwen3.8 family is making "one-prompt tool delivery" real — for enterprises, calling Qwen3.8-Flash, Qwen3.8-27B and Omni-Flash via Alibaba Cloud Model Studio / Bailian is the most cost-effective way to try this generation of models.
📰 September 8, 2026 update: Qwen3.8-27B quantization benchmark
This week a Hacker News benchmark post by Quesma CTO Piotr Migdał gave a definitive answer for running Qwen3.8-27B locally: the 17GB Q4_K_M 4-bit quant matches the 55GB BF16 full model on Terminal-Bench 2.1 and fits on a 24GB consumer card (e.g. RTX 4090) with room for ~64K tokens of context. At 4-bit, GPQA Diamond, IFBench and Terminal-Bench 2.1 scores are statistically indistinguishable from the reference, at one-third the on-disk size. Compression hits a clear cliff at 1-bit: the 6.2GB UD-IQ1_S falls to near-random on GPQA Diamond, and the xhigh reasoning effort actually makes it worse (the model often runs out of token budget without committing to an answer). Unsloth's stated minimum for agentic / tool-calling work remains the 9.8GB Q2_K_XL. For teams weighing a self-hosted 27B deployment, this is the most current, third-party-validated hardware and quantization guidance available.
Source: Quesma: Benchmarking Qwen3.8 27B quantizations (202 HN points, 99 comments on Sept 8, 2026). The original pricing and architecture analysis above remain valid; this card only adds the local-deployment sizing reference.
❓ FAQ
Qwen3.8-Flash vs Qwen3.8-27B — what's the difference?
Qwen3.8-Flash is a MoE model (125B total / 6B active) optimized for cost efficiency and high throughput. Qwen3.8-27B is a dense vision-language model (all 27B active) for deep reasoning at $0.42/$3.08 per million tokens. Choose Flash for scale, 27B for depth.
How does the cache hit pricing work?
When a request hits the context cache (e.g., repeated system prompts, long document prefixes), cached input tokens are billed at $0.016/M instead of $0.15/M — a 90% discount. This dramatically reduces costs for agents, RAG, and multi-turn conversations.
What input modalities are supported?
Qwen3.8-Flash supports text and image input natively (input_modalities: ["text", "image"]), with video analysis in the production version. It handles chart analysis, document OCR, and visual math.
What are the rate limits?
5,000 RPM (requests per minute) and 5,000,000 TPM (tokens per minute), suitable for enterprise-grade high-concurrency deployments.
Start Building with Qwen3.8-Flash Today
$200 free credit + ECS from $4.50/month + open-weight models available
Claim $200 Free Credit →
View ECS Plans
⏰ Alibaba Cloud September Deals · Exclusive Channel
New users: free trial credits + coupon bundle across ECS, databases and more
Model Studio (Bailian) LLM platform: up to 1M free tokens per model for 90 days; up to 50% off for selected models during 22:00-08:00 (UTC+8) off-peak hours
Overseas regions: Singapore / Hong Kong / US ECS with no ICP filing required
-
Enterprise: annual ECS and GPU instances with up to 40% off, plus channel-only pricing on request
🔥 Claim Free Credits →
View ECS Plans →
Offers subject to official campaign pages; new-user deals require identity verification.
© 2026 Lieke Tech · About · Privacy · Contact
Pricing sources: Alibaba Cloud Bailian announcement (Aug 27, 2026), Qwen official blog, Alibaba Cloud Model Studio pricing page
Top comments (0)