guide · ai

Qwen3.7 Flash

Qwen3.7 Flash (Alibaba / Qwen, July 27, 2026): what the release is, why it matters for operators, specs, benchmarks, and the call.

August 12, 2026 · By Alastair Fraser

A retro robot representing Qwen3.7 Flash

--- title: “Qwen3.7 Flash” description: “Qwen3.7 Flash (Alibaba / Qwen, July 27, 2026): what the release is, why it matters for operators, specs, benchmarks, and the call.” type: guide category: ai pubDate: 2026-08-12 image: /images/abs-model-qwen3.7-flash.png imageAlt: “A retro robot representing Qwen3.7 Flash” imagePrompt: “Bold graphic editorial illustration, 1990s comic-book influence, heavy ink outlines, halftone texture, crimson and electric blue on cream. A single retro-futurist robot representing an AI model / a brain-in-a-server, no text, no logos, 16:9.” affiliate: false sources: - name: “eesel AI — Qwen 3.7 Flash review (Aug 6, 2026)” url: “https://www.eesel.ai/blog/qwen-3-7-flash-review” - name: “BenchLM.ai — Qwen3.7 Flash card (Aug 12, 2026)” url: “https://benchlm.ai/models/qwen3-7-flash” - name: “OpenRouter — Qwen3.7 Flash (Jul 27, 2026)” url: “https://openrouter.ai/qwen/qwen3.7-flash” - name: “Vercel AI Gateway — Qwen 3.7 Flash (Jul 28, 2026)” url: “https://vercel.com/ai-gateway/models/qwen3.7-flash” - name: “Roboflow — vision evals panel” url: “https://roboflow.com” - name: “Alibaba Model Studio — Qwen3.7 Flash card” url: “https://modelstudio.aliyun.com” facts: - label: “Vendor” value: “Alibaba / Qwen” - label: “Released” value: “July 27, 2026” - label: “License” value: “Closed-weight (proprietary)” - label: “Type” value: “AI model release” related: - abs-model-claude-opus-5 - abs-model-gpt-5.6-cyber - abs-model-grok-4.6 - abs-model-kimi-k3 - abs-model-lfm2.5-vl-3b - abs-model-ling-3.0-flash - abs-model-muse-glimmer-30b - abs-model-muse-spark-1.2 - abs-model-nemotron-3.5-lightning - abs-model-qwen-image-3.0 - abs-model-qwen3.8-2.4t-a95b - abs-model-qwen3.8-max-weights-update - abs-model-solar-pro-4 tags: - qwen3.7-flash - ai-models - model-release draft: false --- --- title: “Qwen3.7 Flash” description: “Qwen3.7 Flash (Alibaba / Qwen, July 27, 2026): what the release is, why it matters for operators, specs, benchmarks, and the call.” type: guide category: ai pubDate: 2026-08-12 image: /images/abs-model-qwen3.7-flash.png imageAlt: “A retro robot representing Qwen3.7 Flash” imagePrompt: “Bold graphic editorial illustration, 1990s comic-book influence, heavy ink outlines, halftone texture, crimson and electric blue on cream. A single retro-futurist robot representing an AI model / a brain-in-a-server, no text, no logos, 16:9.” affiliate: false sources: - name: “eesel AI — Qwen 3.7 Flash review (Aug 6, 2026)” url: “https://www.eesel.ai/blog/qwen-3-7-flash-review” - name: “BenchLM.ai — Qwen3.7 Flash card (Aug 12, 2026)” url: “https://benchlm.ai/models/qwen3-7-flash” - name: “OpenRouter — Qwen3.7 Flash (Jul 27, 2026)” url: “https://openrouter.ai/qwen/qwen3.7-flash” - name: “Vercel AI Gateway — Qwen 3.7 Flash (Jul 28, 2026)” url: “https://vercel.com/ai-gateway/models/qwen3.7-flash” - name: “Roboflow — vision evals panel” url: “https://roboflow.com” - name: “Alibaba Model Studio — Qwen3.7 Flash card” url: “https://modelstudio.aliyun.com” facts: - label: “Vendor” value: “Alibaba / Qwen” - label: “Released” value: “July 27, 2026” - label: “License” value: “Closed-weight (proprietary)” - label: “Type” value: “AI model release” related: - abs-model-claude-opus-5 - abs-model-gpt-5.6-cyber - abs-model-grok-4.6 - abs-model-kimi-k3 - abs-model-lfm2.5-vl-3b - abs-model-ling-3.0-flash - abs-model-muse-glimmer-30b - abs-model-muse-spark-1.2 - abs-model-nemotron-3.5-lightning - abs-model-qwen-image-3.0 - abs-model-qwen3.8-2.4t-a95b - abs-model-qwen3.8-max-weights-update - abs-model-solar-pro-4 tags: - qwen3.7-flash - ai-models - model-release draft: false --- --- type: guide category: ai model: qwen3.7-flash vendor: Alibaba (Alibaba Cloud / Qwen) release_date: 2026-07-27 status: current license: Proprietary / closed-weight (no weights published, no self-host) --- # Qwen3.7 Flash — Alibaba’s closed-weight “cheap vision” model, and the end of Qwen’s open-weights promise ## The release Qwen3.7 Flash went live on OpenRouter as qwen/qwen3.7-flash on July 27, 2026 (snapshot qwen3.7-flash-20260727), after a one-line QwenCloud changelog entry on July 25 and a Vercel AI Gateway appearance on July 28. BenchLM: “OpenRouter lists qwen/qwen3.7-flash as released on July 27, 2026. QwenCloud’s first-party documentation confirms text, image, and video input, text output, a 1,000,000-token context window, up to 65,536 output tokens, a 256,000-token thinking budget, function calling, built-in tools, and structured output.” The catch: no Qwen blog post, no Twitter announcement, no Hugging Face model card. And the structural fact that matters is Qwen3.7 Flash is proprietary, closed-weight. Older Qwen3.6-Flash was open-weights; 3.7 Flash is not, and neither are Plus or Max. Open weights are no longer a Qwen promise. Flash is the family’s budget/speed-tier sibling below Qwen3.7 Plus (12 tps) and Qwen3.7 Max (text-only flagship, $5.07/M output) — yet the only Qwen3.7 variant that accepts video. In this family, the budget tier sees the most modalities. ## Why it matters for operators The pitch is a vision-language model at $0.03/M input, $0.13/M output that takes text, image, and video in — at ~$0.0001 per sample on Roboflow’s panel, the cheapest vision model on the market and the only sub-$0.10/M model that ingests video. For “tag ten million images by category” workloads, that price changes what is even worth processing. But the headline price is a trap. The $0.03/$0.13 rate applies only under 32K input tokens. Past 256K you land in the top bracket at $0.20/$0.80 per 1M — 6.4× the headline. The 1M context and the cheap price are mutually exclusive. eesel nails it: “So the marketing pairs ‘cheapest vision model’ with ‘1M context’ and lets you assume you get both at once. You do not. The 1M context and the $0.03 price are mutually exclusive. Anything that actually needs a long window lands in the top bracket at 6.7x the rate you budgeted for, whether that is a big codebase or a video or a stack of documents.” OpenRouter’s 30-day blend is $0.044 input / $0.149 output at 51% cache-hit — real traffic pays above list. “Flash” here also means cheap, not fast: 59 tokens/sec, 6× slower than Gemini 3.5 Flash-Lite, P99 90.19s, 8.88% tool-call error rate, one provider (Alibaba Cloud) with no fallback. ## Specs that matter | Spec | Value | |---|---| | Architecture | Not disclosed (dense vs MoE unknown; parameter count unpublished) | | Parameters | Undisclosed | | Context window | 1,000,000 tokens (usable input capped at 991,808) | | Max output | 65,536 tokens (half of Plus/Max) | | Thinking budget | 256,000 tokens (switchable, reasoning on by default) | | Input modalities | Text, image, video (no audio in/out) | | Output modality | Text only | | Per-request media cap | Up to 256 image URLs, 250 Base64 images, or 64 videos | | License | Proprietary / closed-weight (no self-host possible) | | Built-in tools | Code interpreter, web extraction, image search, text search | | Function calling | Yes | | Structured outputs | Documented by Alibaba, not exposed by OpenRouter (only response_format) | | Prompt caching | Yes (Alibaba + OpenRouter) | | Rate limits (Beijing) | 30K RPM / 5M TPM for ≤256K-input tier | ## Benchmarks Qwen published none. No MMLU, no SWE-Bench, no eval table, no parameter count. Artificial Analysis has no model page; LMArena has no Flash entry. There is no intelligence index, no Elo, no published quality score outside Roboflow’s vision-only panel. Independent — Roboflow vision evals (the only public eval that includes it): | Task | Score | Rank (of 23) | |---|---|---| | Overall | 61.7% | #22 | | Identification | 84.4% | #10 | | OCR transcription | 84.1% | #23 (last) | | Object detection | 42.8% mAP@50 | #16 | | Object counting | 46% exact match | Not ranked | | Cost per sample | ~$0.0001 | #1 | eesel: “It lands 22nd of 23 overall at 61.7%, just behind GPT-5.4 mini at 63.5%, and first of 23 on cost. … It can tell you what is in the picture far better than it can tell you where. If your workload is ‘tag ten million images by category’, that is a superb fit at a tenth of a cent per thousand samples. If it is ‘read this invoice and extract the line items’, reach for something else.” OpenRouter telemetry: 59 tps (P50), TTFT 1.20s, end-to-end P50 3.56s / P95 29.03s / P99 90.19s, tool-call error 8.88%, single provider, no fallback. ## Strengths - Sub-cent bulk-vision pricing. At ~$0.0001/sample it changes what is worth processing; 10M input + 2M output runs ~$0.56/month under 32K, ~$3.60/month even at the 1M tier. - Multimodal in, all three modalities. Text + image + video at the cheapest rate in the comparison set, and the only sub-$0.10/M model that takes video. - Genuinely large ceilings. 1M context + 256K thinking budget + 65K output, matched only by Gemini 3.5/3.6 Flash in this tier. - Production-ready surface. 30K RPM / 5M TPM (Beijing), function calling, code interpreter, web/text/image search, prompt caching, OpenAI-compatible API on OpenRouter. - Identification accuracy is usable. 84.4% (#10 of 23) is good enough for robust triage. ## Weaknesses - Closed weights, closed posture. No HF card, no self-host, no architecture disclosure, no parameter count. eesel: “The community has noticed. … For a company that built its reputation on open releases this is a different posture.” LocalLLaMA hoped for open weights that have not materialised. - The pricing bracket trap. Headline $0.03/$0.13 applies only under 32K input; 1M-context work jumps 6.4×, and real traffic already runs above list. - 22nd of 23 on vision, last on OCR. 42.8% mAP@50 detection and 46% exact-match counting make it a poor fit for document extraction or precise localisation. - 59 tps is slow for a “Flash” model. P99 of 90s is fatal inside agent loops. - 8.88% tool-call error rate, single provider, no fallback routing. - Conflicting throughput claims. Ahmad Awais posted 150 TPS in Command Code — contradicting OpenRouter’s measured 59 tps. Trust the independent panel; verify the provider before betting