Qwen3.8-Flash-Next: the Qwen4 preview that changes open-weight cost math

What is Qwen3.8-Flash-Next?

Qwen3.8-Flash-Next is the open-weight checkpoint published August 26 on Hugging Face and ModelScope. Qwen3.8-Flash is the production API built on that architecture — 1M default context, built-in tools (web search, web extraction, image/video search, code interpreter) — live on QwenCloud the same day; OpenRouter listed it that evening. Alibaba confirmed the rate August 27, the specs August 28.

It is multimodal — text, image, and video input; text output — with 512 MoE experts (10 routed + 1 shared per token) on hybrid Gated DeltaNet + QSA attention, and 262,144-token native context extensible to 1,000,000 with YaRN.

Contract note: this is open-weight, not open source — qwen-community-1.0, not Apache-2.0. That distinction matters when you embed a model in client deliverables.

Why Qwen3.8-Flash-Next cuts open-weight MoE inference cost

A dense model passes every token through every parameter: a 125B dense model costs 125B parameters of compute per token. In a MoE, only a subset of experts computes. Qwen3.8-Flash-Next activates about 6B of 125B total — roughly 95% sparsity (6/125 = 95.2%; Alibaba says "125B with 6B activated").

Because inference cost tracks active parameters, that sparsity is the story behind the price. Three design choices reinforce it:

• A 51B-parameter n-gram embedding (bigram/trigram lookup) adds capacity at near-zero per-token compute — it lives in host memory, prefetched asynchronously, so it does not occupy GPU memory.
• QSA attention cuts long-context latency at micro-block level — Qwen reports up to 7.6x prefill / 4.9x decode speedup at 1M context (conservative pair; SGLang cites up to 10.2x/6.6x; both vendor-reported).
• Alibaba puts training at roughly one-ninth the cost of Qwen3.7-Plus — why the price exists at all.

Result: open-weight MoE inference cost at a level once reserved for small dense models — with 1M context, vision, and tools attached.

Qwen3.8-Flash-Next cost: the cheapest AI model to run agents?

Typical agent workload: 10K input / 2K output tokens per task, no cache hits:

• Qwen3.8-Flash: $0.0015 + $0.00094 = $0.0024 per task
• DeepSeek V4 Pro (off-peak): $0.0066 + $0.0040 = $0.0106 per task
• Qwen3.8-Max: $0.02 + $0.012 = $0.032 per task
• Claude Opus 5: $0.05 + $0.05 = $0.10 per task
• GPT-5.6 Sol (list): $0.05 + $0.06 = $0.11 per task

At 100,000 tasks a month: ~$244 on Qwen3.8-Flash vs ~$1,056 on DeepSeek V4 Pro off-peak, ~$10,000 on Claude Opus 5, ~$11,000 on GPT-5.6 Sol — before retries and overhead, on a 1M-context model.

Three agent-specific details push the number lower in practice:

1. Cache reads at $0.016/1M are ~9x cheaper than input — and agent loops re-read history and tool output every turn.
2. 1M context cuts orchestration round-trips — hold the codebase or task history in one window instead of paying for multi-hop summarization. Fewer orchestration round-trips is a cost line, not a free win.
3. Tools plus a thinking toggle keep bills predictable: enable_thinking disables thinking, reasoning_effort (xhigh/medium/low) sets the depth — and at 74 tokens/sec (vs 65 average) there is no slow-thinker tax. OpenCode Go's flat-rate plan added qwen3.8-flash on the same basis.

Where Qwen3.8-Flash-Next lands in the model pricing comparison table

Before you rebuild client quotes, see how it stacks against the field — we track the full set on the AI Model Cost per Task 2026 page. Qwen3.8-Flash-Next slots into the cheap-open-weight tier:

Rates verified 2026-08-27/28 via QwenCloud, OpenRouter, and Artificial Analysis; DeepSeek retired V4-Flash on 2026-09-10: the legacy IDs now serve V4.1-Flash at $0.15/$0.60 off-peak with $0.003 cached input (peak exactly 2×), and DeepSeek V4 Pro keeps its own unchanged row ($0.66/$1.98 off-peak) — so the August Flash band is historical. All USD per 1M tokens. Qwen3.8-Flash matches DeepSeek's current Flash input rate ($0.15) and undercuts it on output ($0.47 vs $0.60), and lands at roughly a quarter of DeepSeek V4 Pro's off-peak price — Intelligent Living frames it as matching V4 Pro's coding score (SWE-bench Pro 62.5 vs 55.4, vendor-reported) at about a quarter of the cost.

Open-weight vs API cost 2026: the agency decision

This release sharpens the open-weight vs API cost decision because both routes are now genuinely competitive at this tier.

When the API wins (most agencies, most months). $0.15/$0.47 via QwenCloud or OpenRouter, zero capex, 1M context, tools, thinking control. At agency volume the bill stays small enough that hardware never pays for itself.

When self-hosting wins. Sustained high volume, data residency, or a per-token cost floor. The FP8 checkpoint is 172.78 GiB (BF16 is 335.28 GiB); two DGX Sparks (128 GB unified each, NVLink) can serve it. That's a deployment project, not an Ollama pull — and qwen-community-1.0 is not Apache-2.0, so free-to-download and free-to-redistribute differ.

The shortcut: below roughly tens of billions of tokens a month, take the API and bank the $0.15 floor. Above that, model the FP8 self-host path against the bill — SGLang shipped day-0 support.

Where it fits in the 2026 model-cost landscape

DeepSeek's August 16 hike (up to 371% on V4 Flash output at peak) reopened the cheap tier — then DeepSeek cut the Flash family again on Sept 10, 2026 with V4.1-Flash ($0.15 in / $0.60 out off-peak, $0.003 cached input), so those Aug 16 Flash numbers are historical; Qwen steps in at $0.15/$0.47 with a stronger active-parameter efficiency story. GLM-5.3-Flash (Z.ai, $0.15/$0.50 list, MIT weights) and Tencent Hy4 preview ($0.834/$2.501, Apache 2.0) made 1M context and cheap open weights table stakes; Qwen3.8-Flash is cheaper on output than both at list. The top tier has not moved — Claude Opus 5 and GPT-5.6 Sol are still 30-60x+ the cheap tier on raw tokens.

Independent data is positive but modest: Artificial Analysis scores it 56 on its Intelligence Index vs a 29 median, at about half the median input price ($0.15 vs $0.30). All benchmarks (SWE-bench Pro 62.5, DeepSWE 58.7, GPQA Diamond 91.7) are vendor-reported, one day old, unreproduced as of August 28 — treat as directional; the weights are public, so third-party runs are coming.

What this changes for your model line item

Qwen3.8-Flash belongs on every agency rate card now: list $0.15/$0.47 with cache reads at $0.016, beside DeepSeek V4 Pro and the open-weight stack.

• Your cost basis dropped again. Priced off GPT-5.6 Sol or DeepSeek V4 Pro off-peak, Qwen3.8-Flash changes the model line item by 4-45x on typical shapes. Decide deliberately whether that becomes margin, lower client price, or faster delivery.
• Cache is the number. At $0.016/1M, re-read-heavy agent loops decide the real monthly bill.
• Budget two prices. Today's list and whatever Qwen charges next quarter — same date-sensitivity asterisk as DeepSeek.

Frequently asked questions

Sources

• Hugging Face model card: huggingface.co/Qwen/Qwen3.8-Flash-Next
• Unite.ai: unite.ai
• OrcaRouter: orcarouter.ai
• QwenCloud pricing: qwencloud.com/models/qwen3.8-flash
• OpenRouter qwen3.8-flash: openrouter.ai/qwen/qwen3.8-flash
• Artificial Analysis: artificialanalysis.ai
• Intelligent Living: intelligentliving.co
• SGLang docs: docs.sglang.io
• OpenRouter API: openrouter.ai/api/v1/models

Accuracy note: Pricing per QwenCloud official rate confirmed 2026-08-27 and OpenRouter's live listing, USD per 1M tokens. 125B/6B/~95% sparsity per Alibaba's Aug 28 announcement and the model card (6/125 = 95.2%). QSA speedups and all benchmarks vendor-reported, unreproduced. DeepSeek V4-Flash is retired as of 2026-09-10 (its IDs bill at V4.1-Flash rates: $0.15/$0.60 off-peak, $0.003 cached input, peak exactly 2×); DeepSeek V4 Pro rates are unchanged. GLM-5.3-Flash and Tencent Hy4 rates per the site's own pricing posts. Per-task math uses 10K/2K tokens; model your own volume.