Qwen 3p7 Plus

Qwen 3.7 Plus is now live on Fireworks

GLM 5.2 is live! Opus-level intelligence at open-source rates. Pay per token on serverless. Try it today.

Blog

Qwen 3p7 Plus Qwen 3.7 Plus on Fireworks: Run it today.

PUBLISHED 6/12/2026

Table of Contents

Qwen 3.7 Plus on Fireworks: Run it today.

Qwen’s latest flagship is built for agent loops

What's launching

Where do your tokens get processed?

The open-weights lineage, and a new path

Benchmarks, briefly

Ways to start using Qwen 3.7 Plus

Get started today

Table of Contents

Qwen 3.7 Plus on Fireworks: Run it today.

Alibaba has partnered with Fireworks to provide inference for their flagship multimodal workhorse model, Qwen 3.7 Plus. Live now on serverless.

Why is this news? A handful of other services listed Qwen 3.7 within hours of Alibaba's June 1st announcement, and that launch announcement was viewed nearly half a million times. This one is different: Alibaba has partnered directly with Fireworks to host and serve the Qwen 3.7 Plus weights on Fireworks infrastructure. There's as much to say about how it's being served as its stellar benchmarks. Qwen’s latest flagship is built for agent loops

To paraphrase the Qwen team's own framing: 3.7 Plus is not a chat model. It is an agent model. It sees the interface, it thinks, it writes code, and takes actions through both GUI or CLI. Verify results, and repeat the loop. Launch materials include an 11-hour app-development session with 10,000+ lines of code and 1,000+ agent calls. That type of workload shape, with long-running loops full of screenshots, tool calls, and cached context, is exactly the type of workload where a frontier-grade inference stack matters. One practical detail for agent builders: the Qwen team recommends preserving thinking across turns for agentic tasks, and Fireworks supports it. Pass reasoning_content back in your assistant messages and set reasoning_history="preserved" , and the model keeps its reasoning from previous turns. Working examples are in the reasoning guide . There are certainly use cases where a long-horizon workflow is not necessary, and setting to reasoning_history="disabled" will likewise reduce unnecessary tokens. What's launching

Qwen 3.7 Plus is a frontier multimodal model. It understands images as well as text, and supports both thinking and non-thinking modes: request reasoning when the task needs it, skip it when it does not, same as any other hybrid-reasoning model. Qwen 3.7 Plus is now on serverless : pay per token, no infrastructure to manage, no cold starts. The facts are in the card below. Qwen 3.7 Max is coming soon; talk to us if you want early access. Model Specs Model ID accounts/fireworks/models/qwen3p7-plus Modalities Text + image input, text output Reasoning Thinking and non-thinking, selectable per request Context window 262k tokens. Serverless Pricing $0.50 / 1M input · $0.10 / 1M cached input · $3.00 / 1M output Batch 50% below serverless pricing, results within 24h Training limited availability, contact us if interested. Weight export Not available (licensed weights; tuned models serve on Fireworks). APIs OpenAI-compatible and Anthropic-compatible

On pricing: routed listings may quote lower numbers, and that is fine. Fireworks is the path for teams that want Qwen 3.7 Plus on Fireworks-operated infrastructure, under Fireworks data handling and SLAs, and at the speed, latency, and throughput that AI natives have come to expect. Where do your tokens get processed?

There are two ways an API can deliver tokens from Qwen 3.7 Plus: API routers and inference providers . An API router receives your request and forwards it to an inference provider (e.g official Alibaba API), which then runs the inference. Routers are generally useful because they provide convenience: a single API key, many models, fallback options, and consolidated billing. Genuinely useful. An inference provider runs the model directly. Fireworks is an inference provider. When you call our API, your request is processed on our infrastructure by our inference engine using weights we host. We do not forward your traffic to another endpoint. Ever. As an inference provider, we make commitments about latency (we control the full serving path), data governance (e.g our zero data retention policy ), and reliability (our 99.9% uptime SLA). For additional details, we've written a longer piece on the differences between inference providers and API routers, and the implications for each. Hosting closed-weight models is not new territory for us. We have meaningful experience serving proprietary model weights under license for hundreds of enterprise customers like Cursor and Notion . What makes the Qwen 3.7 Plus program relatively novel is that we're making such a model available to all developers in the Fireworks community. Alibaba has licensed Fireworks to host and serve these weights directly, which means you get independent inference infrastructure, not a proxy, with all the performance and data handling properties that entails. The open-weights lineage, and a new path

Qwen earned its position the open way: by HuggingFace download data analyzed by Interconnects , Qwen has become one of the default open model families for local inference, reasoning, and multimodal workloads. Qwen 3.7 is the most capable release in that lineage, on a different distribution path: the Plus weights are not on HuggingFace. They are licensed, and Fireworks serves them independently under that license. We want to be direct about that distinction rather than blur it. If you build on open-weights Qwen models, we serve those too, on the same stack. Qwen’s 8B parameter models remain among the most popular small models for fine-tuning on the Fireworks training platform . Benchmarks, briefly

Benchmark tables are sticker specs: necessary, not sufficient. The full tables are in Qwen's launch post , but three results stand out. Qwen 3.7 Plus with thinking enabled matches the flagship Max on AIME 2025 (14/15) at roughly three times the speed. End-to-end throughput is 3.55× faster than Qwen 3.6 Plus. And on coding-agent tasks, early results show strong generalization across multi-file development, debugging, and issue-resolution workflows. Model-quality benchmarks from Qwen's launch materials. These measure the model, not Fireworks serving. We will publish Fireworks serving benchmarks separately, with methodology. Ways to start using Qwen 3.7...