Qwen3.8-Flash

Model IntroductionText

Qwen3.8-Flash-Next is the first open-weight model built on the architecture that will underpin Qwen4 — an experimental preview centred on efficient scaling.

Introduction

As the frontier of foundation models pushes toward ever-larger parameter counts and ever-longer context windows, the question is no longer just how much we can scale, but how efficiently we can do so. Sustainable progress toward artificial general intelligence (AGI) that benefits everyone demands architectural innovation. Today, we are sharing a concrete step in that direction: Qwen3.8-Flash-Next. This experimental preview of the architecture that will underpin Qwen4 is built around a fundamental rethinking of how the core components of modern large language models (LLMs) interact at scale.

Key Features

  • Hybrid Attention with QSA: The Gated DeltaNet and Gated Attention pairing has been reworked into Gated DeltaNet and Qwen Sparse Attention (QSA). Rather than selecting individual tokens for processing, QSA operates at the micro-block level. This cuts long-context latency significantly, a critical gain as agentic workloads increasingly dominate real-world usage.
  • Gated Residual: Residual streams with normalization are what make deep LLM training manageable. Gated Residual modulates information flowing through widened residual streams via an element-wise, data-dependent read gate and a per-branch scalar write gate. This brings finer-grained expressiveness across layers while preserving training stability and keeping inference overhead low.
  • N-gram Embedding: Embeddings provide a unique axis for parameter scaling that requires less computation and is more amenable to offloading than Mixture-of-Experts (MoE). By indexing with short n-grams, this approach makes parameter scaling highly efficient for memory-constrained accelerators without sacrificing quality.
  • Tailored Training Recipe: The Muon and AdamW optimizers are applied to specific weight categories to maximize efficiency. Guided by refitted scaling laws, we eliminate traditional batch-size warmups and start directly at the target batch size, substantially reducing total optimizer steps while safely supporting larger learning rates for robust convergence.

Benchmarks

Benchmark (Area)Score*
Agentic coding DeepSWE 1.1 (Coding)58.7 (Eval)
Agentic coding SWE-bench Pro (Coding)62.5 (Eval)
Multilingual software engineering SWE-bench Multilingual (Coding)81.0 (Eval)
Repo-level code generation NL2Repo-Bench (Coding)48.1 (Eval)
Long-horizon office work CoWorkBench (Agent)73.9 (Eval)
Professional job tasks JobBench (Agent)55.7 (Eval)
Frontier agentic tasks Agents' Last Exam (Agent)Pass@1 24.3 Score 51.2 (Eval)
Real-world tool use Toolathlon Verified (Pass@1) (Agent)73.5 (Eval)
Instruction following IFBench (General)81.3 (Eval)
Scientific reasoning GPQA Diamond (General)91.7 (Eval)
Multidisciplinary reasoning HLE (General)35.9 (Eval)
Competitive coding LiveCodeBench v6 (General)91.9 (Eval)
Multimodal tool use ClawEval-MM (Agentic Multimodal Intelligence)Pass@3 64.4 Average 60.4 (Eval)
Application recreation RecreationBench (Agentic Multimodal Intelligence)49.9 (Eval)
Mobile use AndroidWorld (Agentic Multimodal Intelligence)84.5 (Eval)
Computer use OSWorld 2.0 (Agentic Multimodal Intelligence)Binary 19.4 Partial 52.3 (Eval)
Visual web development Vision2Web (Agentic Multimodal Intelligence)64.0 (Eval)
Embodied intelligence ERQA (General Multimodal Intelligence)72.3 (Eval)
Long video understanding LVBench (General Multimodal Intelligence)76.6 (Eval)
Real-world perception RealWorldQA (General Multimodal Intelligence)88.5 (Eval)
Visual math problem solving MathVision (General Multimodal Intelligence)Without CI 90.6 With CI 95.7 (Eval)
Scientific chart analysis CharXiv (RQ) (General Multimodal Intelligence)Without CI 84.6 With CI 90.6 (Eval)

API request example

Endpoint /v1/chat/completions. Replace <API_KEY> with a key from the console.

Standard OpenAI Chat Completions. Swap `model` for the target slug.

curl https://api.ravenrouter.ai/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer <API_KEY>" \
  -d '{
  "model": "Qwen3.8-Flash",
  "messages": [
    {
      "role": "user",
      "content": "Hello! Introduce yourself in one sentence."
    }
  ],
  "max_tokens": 64
}'

More models: API Reference → Examples by model

Qwen3.8-Flash API — $0.13/1M input, $0.43/1M output | RavenRouter AI