Qwen3.7 Max
Alibaba's agent-frontier flagship — the fastest frontier model at 171 tokens/sec with a 2.5-second first token, Intelligence Index 57, #1 on instruction following (IFBench 80.5%), and a demonstrated 35-hour autonomous agent run. Released May 19, 2026.
Model Specs
- Released
- May 2026
- Context window
- 1.0M tokens
- Max output
- 66K tokens
- Capabilities
- reasoningfunction-callinglong-contextagentic-workflows
- Modalities
- text
About this model
Qwen3.7 Max is the flagship of Alibaba's Qwen3.7 series, released May 19, 2026 under the banner "The Agent Frontier." It scores 57 on the Artificial Analysis Intelligence Index — #5 overall at launch, the highest-placed Chinese model — while being, remarkably, **the fastest frontier model on the market: 171.5 tokens/sec with a 2.47-second time-to-first-token**. Where Fable 5 and GPT-5.5 deliberate for 100+ seconds before answering, Qwen3.7 Max starts streaming almost immediately.
The agent story is the headline. Alibaba demonstrated a **35-hour autonomous agent run — 1,158 tool calls without degradation** — and the benchmarks back the positioning: #1 of all models on IFBench (80.5%, precise instruction following), 94.7% on τ²-Bench Telecom, 50.8% on Terminal-Bench Hard, and official scores of 80.4% on SWE-bench Verified, 60.6% on SWE-Bench Pro, and 91.6% on LiveCodeBench. GPQA Diamond lands at 92.3% — within two points of every Western flagship. Notably, AA measured the **lowest hallucination rate among frontier models** (22.9%) on AA-Omniscience's incorrect-answer metric.
Two caveats define the trade. First, Qwen3.7 shipped **closed-weights** — a break from Qwen's open-weight tradition (announced 27B/35B open variants haven't shipped). Second, it's text-only: no image/video input, unlike Gemini or the multimodal sibling Qwen3.7-Plus. Pricing is $2.50/M input and $7.50/M output (Alibaba list; some hosts run promos), with a 1M-token context — up from 256K on Qwen3.6. Reach for Qwen3.7 Max when you're building agents that need frontier capability at interactive speed: tool-calling loops, long-horizon automation, and high-throughput production pipelines.
Key Strengths
Fastest frontier model
171.5 tokens/sec with a 2.47-second first token — an order of magnitude more responsive than deliberation-heavy flagships (GPT-5.5 ~103s, Fable 5 ~108s TTFT) at comparable intelligence.
#1 instruction following
80.5% on IFBench — the highest of any model — meaning it does precisely what you specify. The single most important trait for production agents and structured pipelines.
35-hour autonomous agent runs
Alibaba demonstrated 1,158 consecutive tool calls over 35 hours without degradation. τ²-Bench Telecom 94.7% and Terminal-Bench Hard 50.8% confirm the long-horizon agentic design.
Strong real-codebase coding
SWE-bench Verified 80.4%, SWE-Bench Pro 60.6% (ahead of GPT-5.5's 58.6%), and LiveCodeBench 91.6% — frontier-grade engineering at a third of US flagship prices.
Lowest frontier hallucination rate
22.9% incorrect-answer rate on AA-Omniscience — the lowest among frontier models — making its fast answers also unusually trustworthy.
1M context at value pricing
$2.50/M input, $7.50/M output with a million-token window (4x Qwen3.6's 256K) — half of GPT-5.5's input rate and a quarter of its output rate.
How it compares
Qwen3.7 Max competes on speed-adjusted frontier intelligence — near-flagship depth at interactive latency and value pricing.
| vs. Model | Verdict | Outcome |
|---|---|---|
| DeepSeek V4 | The two Chinese frontier leaders: Qwen3.7 Max is smarter (Index 57 vs 52), 4x faster (171 vs 45 t/s), and leads IFBench (80.5 vs 71.3); DeepSeek V4 is ~6x cheaper ($0.435/$0.87), MIT open-weights, and slightly ahead on LiveCodeBench (93.5 vs 91.6). Performance-critical agents → Qwen; budget and self-hosting → DeepSeek. | Wins most cases |
| GPT-5.5 | GPT-5.5 leads composite intelligence (60.2 vs 57) and GDPval professional work, but Qwen3.7 Max beats it on SWE-Bench Pro (60.6 vs 58.6), IFBench (80.5 vs 75.9), and is ~40x faster to first token (2.5s vs ~103s) at half the price. Depth-critical work → GPT-5.5; agents and interactive flows → Qwen. | Depends |
| Gemini 3.1 Pro | Even on the Intelligence Index (both 57) with opposite profiles: Gemini 3.1 Pro wins science (GPQA 94.1 vs 92.3, ARC-AGI-2) and adds video/audio input; Qwen3.7 Max wins agentic work (IFBench, tau2, 35-hour runs) and raw speed (171 vs 110 t/s). Research → Gemini; agents → Qwen. | Depends |
Pros
- Fastest frontier model: 171.5 t/s, 2.47s first token
- #1 of all models on IFBench instruction following (80.5%)
- Demonstrated 35-hour autonomous agent run (1,158 tool calls)
- SWE-Bench Pro 60.6% — ahead of GPT-5.5
- Lowest hallucination rate among frontier models (22.9%)
- 1M context at $2.50/$7.50 — half of GPT-5.5's input rate
- τ²-Bench Telecom 94.7% for production tool-calling
Things to consider
- Closed weights — a break from Qwen's open-weight tradition (27B/35B variants announced but unshipped)
- Text-only: no image/video input (the Plus sibling adds vision)
- Trails Western flagships on composite intelligence (57 vs 60-65)
- Very verbose reasoning internally (~97M tokens on the AA index run)
- Knowledge cutoff not disclosed by Alibaba
- Weaker on hardest physics/science evals (CritPt 13.4%)
Best use cases
Production AI agents
The design target: fast tool-calling loops, precise instruction following, and proven long-horizon stability for agents that run for hours.
Interactive coding assistants
Frontier-grade code quality (SWE-V 80.4%) at 171 t/s — fast enough for IDE-style flows where 100-second waits are unusable.
High-throughput pipelines
Frontier intelligence at speed-tier latency makes batch processing, extraction, and structured generation dramatically cheaper per unit time.
Structured output workflows
#1 IFBench score means JSON schemas, strict formats, and multi-constraint specs are followed with best-in-class precision.
Long-document processing
1M context plus AA-LCR 69% — contracts, codebases, and document sets analyzed in one fast request.
Cost-controlled frontier work
Near-flagship quality at $2.50/$7.50 — the value play between budget models and $5-10 Western flagships.
How to use it on Renas AI
- 1
Step 1
Open AI Chat on Renas
Navigate to AI Chat in the Renas dashboard and pick the most capable Qwen model available to your plan.
- 2
Step 2
Write precise instructions
This is the IFBench champion — detailed constraints, formats, and multi-step specs are followed exactly. The more precise your spec, the more the model shines.
- 3
Step 3
Exploit the speed
2.5-second first tokens mean you can iterate conversationally — treat it like a real-time pair programmer rather than a batch oracle.
- 4
Step 4
Chain tool-heavy workflows
Long-horizon stability is the differentiator: multi-step agent flows with many tool calls hold together where other models drift.
Pricing
Pricing on Renas AI
Pay-as-you-go credits, no API keys, no rate limits.
~200,000 words on a 10,000-credit Spark plan
Frequently asked questions
Other text models on Renas AI
The agent frontier on Renas
Access leading Qwen models with your Renas AI subscription credits — no API key, no setup, no per-seat fees.
Try Qwen on Renas