Back to feed
News Story
APriority71
The Batch (Andrew Ng)
1 sources

Issue 356

Original

No summary is available for this story yet.

SynthePulse Insight · AI deep reading

Qwen3.7-Max Challenges Google, Joins Top Tier of Large Models

Version 1 · 1 source

Alibaba releases Qwen3.7-Max, approaching top US models in multiple benchmarks, but its closed-source strategy and unverified agent capabilities raise concerns.

  • Qwen3.7-Max ranks seventh in the Artificial Analysis Intelligence Index, trailing top models from OpenAI, Anthropic, and Google.
  • The model achieves the lowest hallucination rate (23%) on the AA-Omniscience factual knowledge test, but by refusing to answer more than half of the prompts.
  • Output speed reaches 208 tokens/second, tying with Gemini 3.5 Flash for third place.
  • Alibaba claims agent capabilities based on internal tests, not yet verified by independent benchmarks.
  • Model weights are not open-sourced, continuing Alibaba's closed-source strategy since late 2025.
  • Pricing is $2.50/$0.25/$7.50 per million input/cached/output tokens.
Open section navigationPerformance: Approaching Top Tier, but Still Gaps

Performance: Approaching Top Tier, but Still Gaps

Qwen3.7-Max scores 56.6 (reasoning mode) on the Artificial Analysis Intelligence Index (aggregating 10 economically useful tasks), ranking fifth or seventh depending on other models' reasoning levels. It trails Gemini 3.1 Pro Preview (57.2) but leads Gemini 3.5 Flash (55.3). The evaluation consumed approximately 97 million output tokens, far above the average of 35 million.

On the AA-Omniscience factual knowledge test, Qwen3.7-Max ranks sixth (14 points), far below Gemini 3.1 Pro Preview (33 points) but above Claude Sonnet 4.6 (12 points). Its 23% hallucination rate is the lowest among frontier models, but achieved by refusing to answer more than half of the prompts.

Agent Capabilities: Internal Tests, Awaiting Verification

Alibaba emphasizes Qwen3.7-Max's agent capabilities, but this claim is based on internal tests not yet verified by independent benchmarks. The model reportedly autonomously optimized an attention kernel over 35 hours, calling 1,158 tools and running 432 kernel evaluations.

Alibaba's reinforcement learning method separates training of tasks, agent tool sets, and verifiers to prevent the model from learning tricks specific to a setup. However, these details come from Alibaba's description and lack third-party confirmation.

Closed-Source Strategy and Pricing

Like other top Qwen models since late 2025, Qwen3.7-Max's weights are not open-sourced. Alibaba also released the multimodal model Qwen3.7-Plus-Preview.

API pricing is $2.50/$0.25/$7.50 per million input/cached/output tokens, with a free version available via Qwen Chat (account registration required). Model parameter count, architecture, training data, and methods are not disclosed.

Credibility boundary

This article is primarily based on reporting from The Batch (Andrew Ng), a secondary source that cites third-party benchmarks such as Artificial Analysis. Alibaba's internal test claims have not been independently verified and should be treated as source statements.

Insight takeaway

Qwen3.7-Max approaches top US models in performance, particularly in speed and low hallucination rate, but its closed-source strategy and unverified agent capabilities mean it has not yet fully established parity with OpenAI, Anthropic, and Google.

Sources for this version

  1. Issue 356The Batch (Andrew Ng)

Primary report

The Batch (Andrew Ng)

Primary source