Qwen 3.8 Flash Is Now on Multos — A 125B Model That Costs Less Than DeepSeek V4 Flash

Alibaba released Qwen 3.8 Flash on August 26, 2026, and the headline number is a little hard to believe at first: it beats DeepSeek V4 Pro on SWE-bench Pro — 62.5 versus 55.4 — while costing roughly a quarter of the price. It activates only 6 billion parameters per token out of a total of 125 billion, making it one of the most efficient models available right now. We've added it to Multos via Novita AI at the Lite tier, same cost band as GLM-5.3 Flash and DeepSeek V4 Flash.

How a 125B Model Activates Only 6B Parameters

Qwen 3.8 Flash is a mixture-of-experts (MoE) model. Instead of running all 125 billion parameters on every token, it routes each token through a small subset of specialist expert networks — 10 routed experts plus 1 shared, out of 512 total. Three out of every four attention layers use Gated DeltaNet (GDN), a linear-attention mechanism that compresses historical context into a fixed-size state rather than recomputing full attention over the entire sequence. This is why the model handles a 1 million token context without the compute cost exploding — at 1M tokens it achieves up to 7.6x faster prefill than a standard transformer. On top of that there are 51 billion N-gram Embedding parameters stored as a lookup table for common phrases and patterns, living in host memory and prefetched asynchronously during inference. More capacity without more per-token compute — a genuinely new approach to scaling.

The Benchmarks, Honestly

The SWE-bench Pro number is the one people are talking about: 62.5 with 6 billion active parameters, beating DeepSeek V4 Pro's 55.4 with 49 billion active parameters. That's a structural argument that smarter architecture beats brute-force parameter count. On office and professional tasks it also leads: CoWorkBench 73.9 versus 45.1 for DeepSeek V4 Flash, JobBench 55.7 versus 41.3. Where it doesn't win: frontier reasoning (HLE) where DeepSeek V4 Pro scores 42.7 versus 35.9, and DeepSWE 1.1 where V4 Pro edges ahead at 62.7 versus 58.7. Claude Opus 5 remains the coding leader at 79.2 on SWE-bench Pro, but at $5 per million input tokens versus $0.15 for Qwen 3.8 Flash — you're getting 79% of Opus 5's coding performance at 3% of the price.

Why It's at the Lite Tier Despite Strong Numbers

Qwen 3.8 Flash's pricing on Novita AI puts it firmly in the flash/lite cost band. The reason it lands at Lite rather than Standard is straightforward: it's cheaper to serve because the active parameter count is 6B per token versus 18-49B for standard models, and we price by cost band rather than just capability. The native context window is 262,144 tokens, extending to 1 million via YaRN interpolation. It processes text, images, and video. Implicit caching activates automatically on repeated prefixes at $0.016 per million tokens — no setup required, no explicit cache management needed.

What You Can Use It For on Multos

Coding tasks where you're cost-sensitive: SWE-bench Pro 62.5 is genuinely good for fixing bugs, writing functions, and reviewing pull requests at a fraction of what you'd spend on a Pro-tier model. Long agentic workflows: the hybrid attention architecture was specifically designed for long-context agentic use — the model doesn't slow down as context grows. Office and professional tasks: CoWorkBench and JobBench scores suggest it handles structured knowledge work well — document analysis, spreadsheet reasoning, multi-step research tasks. Vision and code combined: it processes images and video natively, so drop a screenshot of an error or a UI mockup into your conversation. High-volume production: at $0.15 per million input tokens with automatic implicit caching at $0.016, running at scale is cheaper than most alternatives.

An Architecture Preview of Qwen 4

Qwen 3.8 Flash is explicitly an architectural preview of Qwen 4. Alibaba released four design changes — hybrid GDN and sparse attention, Gated Residual, N-gram Embedding, and the Muon optimizer — early so the community can evaluate them before the full Qwen 4 family arrives. Just as Qwen3-Next previewed the architecture used across the Qwen3.5 through Qwen3.8 series, this model introduces the design that will underpin Qwen 4. Getting familiar with it now means you're already ahead of the curve when the next generation lands. The model required only about one-ninth the training compute of Qwen3.7-Plus — if these efficiency gains hold at scale, they suggest the economics of AI development are shifting faster than most people expect.

Ready to Get Started?

Qwen 3.8 Flash is live on Multos Lite and above via Novita AI. 125B total parameters, 6B active per token, 1M token context, native multimodal input, $0.15 per million tokens. Select it under Novita AI in the model dropdown or let Auto Router pick it — available from the $20/month Lite plan upward.

Start Building Free