Alibaba drops Qwen3.8-Flash: 125B open-weight MoE with 6B active params

Alibaba just rolled out Qwen3.8-Flash, an open-weight multimodal MoE model built for high throughput, heavy tool calling, and coding agents without the massive compute bill.

It packs a 125B-parameter main model alongside 51B N-gram embeddings, yet it only activates 6B parameters per token. It handles 262K context natively, stretches up to 1 million tokens, and serves as an architectural preview of the upcoming Qwen4 series.

What it can do

Qwen3.8-Flash targets real-world production tasks where cost and latency usually kill agent workflows.

On benchmarks, it goes toe-to-toe with DeepSeek-V4-Flash and Claude-Opus-4.6. It puts up strong numbers on agentic coding in SWE-bench Pro, office workflows in CoWorkBench, and real-world tool use in Toolathlon Verified.

On the multimodal front, it holds its own on visual math with MathVision, mobile device control on AndroidWorld, and embodied reasoning on ERQA.

Alibaba already put it into production on its QwenWork platform. Powering the redesigned Standard mode, the model slashed token consumption per task by 75% while roughly doubling generation speed.

Under the hood and technical specs

The efficiency jump comes from major plumbing changes across attention and layer connections.

It runs a hybrid attention setup pairing Gated DeltaNet, which compresses historical context, with Qwen Sparse Attention to pick out relevant tokens through a lightweight compressed indexer. This keeps attention costs flat over massive context windows.

It also uses Gated Residuals to strengthen cross-layer signal flow and stability, scales capacity cheaply through 51B N-gram embeddings, and relies on the Muon optimizer during pretraining.

Because of these changes, Qwen3.8-Flash required just one-ninth of the training compute used by Qwen3.7-Plus—a model three times its physical size—while beating it on code and enterprise workflows.

How to get it

The weights are fully available for download today on Hugging Face and ModelScope under open access.

If you prefer cloud endpoints, Alibaba has hooked it up to Model Studio and Qwen Cloud.

Pricing is set aggressively at $0.16 per million tokens for input and $0.47 per million tokens for output.

Why it matters

If you are running long-running agents or high-volume code assistants, this is an instant test candidate. You get near-frontier coding and tool calling while paying the compute footprint of a tiny 6B model.

Beyond immediate cost cuts, this drop shows where open weights are heading. Pairing sparse attention with massive sparse parameter MoEs is Alibaba's blueprint for Qwen4, and it proves high-end reasoning doesn't require brute-force inference budgets anymore.