The Qwen Team has released Qwen3.8-Flash-Next, an open-weight multimodal mixture-of-experts model that serves as an early architectural preview of the upcoming Qwen4 series. The model combines substantial capacity with strong cost efficiency, with 125 billion total parameters and an additional 51 billion N-gram embedding parameters, while activating only 6 billion parameters per token.

The release follows Qwen3-Next, which introduced hybrid architecture designs that were later adopted across the Qwen3.5 through Qwen3.8 families. Qwen3.8-Flash-Next will be available through the QwenCloud API at $0.16 per million input tokens and $0.47 per million output tokens, positioning it for workloads where both throughput and token costs matter, including coding assistants and enterprise agentic workflows.

Benchmark results suggest strong performance across software engineering and autonomous agent tasks. The model scored 58.7% on DeepSWE 1.1, 62.5% on SWE-bench Pro, and 81.0% on the multilingual variant. In long-horizon office automation measured by CoWorkBench, it scored 73.9%, ahead of both Qwen3.7-Plus and Claude-Opus-4.6.

Its general reasoning performance was also strong, with scores of 91.7% on GPQA Diamond and 91.9% on LiveCodeBench v6. Multimodal performance remained solid as well, including 84.5% on AndroidWorld and 76.6% on LVBench for long video understanding. Native context length reaches 262,144 tokens and can be extended to one million tokens via YaRN, which makes the model relevant for longer documents and multi-step workflows that need larger context windows.

Meet Qwen3.8-Flash, a multimodal MoE and an early preview of the Qwen4 architecture, now open-weight! The production version Qwen3.8-Flash will be available soon via QwenCloud API at just $0.16/1M input tokens and $0.47/1M output tokens. 125B parameters + 51B N-gram… pic.twitter.com/SScnmzWS7O — Qwen (@Alibaba_Qwen) August 26, 2026

Architectural Innovations and Developer Integration

The architecture introduces four system-level upgrades. For attention, the model combines Gated DeltaNet with Qwen Sparse Attention, compressing historical context efficiently while retrieving relevant information through micro-block indexing rather than token-level processing. According to the release, this design delivers up to 7.6x prefill speedup at one million tokens.

The Gated Residual mechanism expands the residual stream into four parallel branches with dynamic gating. The Qwen Team says this improves cross-layer information flow and training stability while supporting FP8 storage to reduce memory traffic. N-gram Embedding adds capacity through local-context lookups that require minimal computation and can be asynchronously prefetched from host memory.

Training uses the Muon optimizer with refined orthogonalization and parameter-splitting strategies, which the team says enables stable convergence at larger batch sizes without traditional warmup procedures.

The model weights are available on HuggingFace and ModelScope. API access is provided through QwenCloud, with support for OpenAI-compatible Chat Completions and Anthropic-compatible protocols. Developers can integrate the model into existing workflows through Claude Code, OpenAI Codex, Qoder CLI, Qwen Code, and OpenClaw, with reasoning effort configurable across low, medium, and xhigh levels. The team said an official production release with built-in tools and a default one-million-token context is expected to follow shortly.