Alibaba's Qwen3.8-Flash Launches: New Architecture Cuts Costs 9x to Beat Claude
AGI Hunt · wechat · 2026-08-28
Alibaba Qwen released Qwen3.8-Flash-Next, featuring the new Next architecture as a prototype for the Qwen4 series.
Core Architecture & Performance
- MoE & Sparse Activation: A 125B parameter main model with a 51B N-gram Embedding addon activates only 6B parameters per token, yet outperforms Claude Opus 4.6 and approaches Opus 4.8.
- Long Context Optimization: Combines GatedDeltaNet and QwenSparseAttention (QSA), achieving up to 7.6x speedup on 1M token contexts while significantly reducing costs.
- 4-Lane Residual Streams: Uses four parallel branches with dynamic gating, where one branch spontaneously evolved into a long-range pathway for better cross-layer information flow.
- Offloaded Memory: The N-gram Embedding acts as a dictionary for common phrases stored in host memory, saving GPU VRAM with minimal compute cost.
Benchmarks & Pricing
Excels in multimodal tasks (image recognition, SVG generation) and coding. API pricing is extremely low at 0.8 RMB (input) and 2.7 RMB (output) per million tokens—less than 1/40th of Claude Opus 4.6. Training costs were reduced to 1/9th of the previous generation. Inference frameworks like vLLM and SGLang support it on day one, with various quantized versions available.
More from Models
- "Argon doesn't trust me": researcher laments overzealous model refusal — thegautamkamath · 2026-10-15
- User finds Grok shockingly bad at finding X posts, prefers Opus with browser — JasonBotterill · 2026-10-15
- "Claude-Next" Reportedly Returns to Claude CLI and Web via Stealth Opus Routing — koltregaskes · 2026-10-15
- 'Thinking leakage': no-think RL gains in hybrid models just reuse think-mode skills — kaust-generative-ai · 2026-10-15
- Cline: Gemini 4 Argon saw 3x Claude's day-one adoption and 30% fewer tokens — JMateosGarcia · 2026-10-15
- Same prompt, three AIs design a pro UI for a JUCE shimmer reverb — Grok 4.7 ties Opus 5.5 — hullabaloo22 · 2026-10-15