Qwen3.7-Max: The Agent Frontier (Qwen)

TL;DR

Qwen3.7-Max is Alibaba’s new closed-weight flagship reasoning/agent model, formally announced at the 2026 Alibaba Cloud Summit on May 20 after a quiet LM Arena preview a few days earlier. It targets the long-horizon-agent regime: a 1M-token context window (up from 256K on Qwen3.6-Max) and headline numbers of ~35 hours of autonomous task execution with 1,000+ tool calls in a single run without performance collapse. On Artificial Analysis Intelligence Index v4.0 it scores 56.6, ranking #5 overall and the top Chinese model; on LM Arena it sits at Elo ~1,475 (#13 in text). A multimodal Qwen3.7-Plus-Preview is also previewed; Max is text-only.

Key claims

Qwen3.7-Max is positioned as an “agent-frontier” reasoning model with a 1M-token context window (vs 256K on Qwen3.6-Max), targeting sustained autonomous execution rather than single-turn quality [§The Agent Frontier].
Internal testing reports ~35 hours of autonomous task execution with 1,000+ tool calls in a single run without performance falling off [§Agentic capability].
On Artificial Analysis Intelligence Index v4.0 the model scores 56.6, a 4.8-point gain over Qwen3.6-Max-Preview and #5 overall (top Chinese model) [§Benchmarks].
On LM Arena Text Arena the preview entered at Elo ~1,475 (#13 overall; top-10 in math and coding) — a “leaderboard-topping composite score, mid-pack human-preference” split worth flagging [§Arena].
Release is closed-weight, API-only: available on chat.qwen.ai and via Alibaba Cloud Model Studio / DashScope, with OpenRouter pricing at $2.50 /$ 7.50 per 1M input / output tokens; Plus tier may follow earlier patterns and open-source, Max stays proprietary [§Availability].
API surface advertises both OpenAI- and Anthropic-protocol compatibility, including drop-in use with Claude Code [§API compatibility].
A multimodal companion preview (Qwen3.7-Plus-Preview) is also available and ranked #16 in LM Arena Vision; the Max tier itself is text-only with no image input [§Plus preview].

Method

The qwen.ai blog page is JS-rendered and did not return body text to the fetcher; the claims above are reconstructed from the qwen.ai/blog announcement title (“Qwen3.7: The Agent Frontier”), Alibaba’s official @Alibaba_Qwen tweet (May 18), and contemporaneous independent coverage (MarkTechPost, Decrypt, IndexBox, multiple Codersera/Aimadetools writeups). Per BUD.md §8 the page intentionally avoids architectural claims the announcement does not substantiate — there is no parameter count, no training-data description, no RL recipe, and no public technical report at filing time. What the announcement does commit to is a capability profile (sustained autonomous tool-use at 1M-token context, with a Max/Plus tier split) and a distribution strategy (closed weights, multi-protocol API).

The “agent frontier” framing has a specific operational shape: the model is shipped locked in deep-thinking mode for the preview (chain-of-thought visible, web search and code interpreter disabled in chat.qwen.ai), and the headline result is a hand-picked single-run demo of ~35 h / 1,158 tool calls executed by Qwen3.7-Max with the SGLang Triton kernel as a serving substrate.

Results

Reported numbers, as of the May 20–21 announcement window:

Intelligence Index v4.0: 56.6 (#5 overall, +4.8 over Qwen3.6-Max-Preview). Behind GPT-5.5 (60.2), Claude Opus 4.7 (57.3), Gemini 3.1 Pro Preview (57.2); ahead of Gemini 3.5 Flash (55.3).
Context window: 1,000,000 tokens (4× over Qwen3.6-Max-Preview’s 256K). No independent long-context retrieval evaluation at filing time.
LM Arena Text: Elo ~1,475, #13 overall; top-10 in math and coding sub-arenas. Alibaba reports #6 lab in Text, #5 in Vision after the preview drop.
Terminal-Bench Hard: 50.8% (third-party summary; not independently reproduced).
Apex Math: 44.5 (third-party summary, vs Claude Opus 4.6 Max at 34.5).
Hallucination rate: 22.9% — claimed lowest among frontier models in the same third-party summary; AA-Omniscience separately shows raw accuracy down with abstention up vs Qwen3.6, consistent with an abstain-when-unsure tuning.
Long-horizon execution: 1,158 tool calls in 34.7 hours in Alibaba’s internal demo; no independent reproduction.
Pricing: $2.50 /$ 7.50 per 1M input/output tokens on OpenRouter (qwen/qwen3.7-max); compares to Claude Opus 4.7 at $15/$ 75 and GPT-5.5 at $10/$ 30.

Why it’s interesting

This is the second filed datapoint (alongside Kimi K2.5: Visual Agentic Intelligence) of a Chinese-lab frontier model claiming a long-horizon agent axis as its primary differentiator rather than per-turn benchmark scores — and the first to put a concrete number on it (1,158 tool calls / 34.7 h). It also extends the “leading Qwen models go closed” trajectory previewed at Qwen3.6-27B announcement: 27B dense beats Qwen3.5-397B-A17B on coding benchmarks for Qwen3.6-27B (open) vs Qwen3.6-Max (closed), and continues the closed-but-API-accessible cohort that Open foundation-model releases tracks as the comparison baseline for open releases. The 1M-token / sustained-tool-use framing complements MiroThinker: Pushing the Performance Boundaries of Open-Source Research Agents via Model, Context, and Interactive Scaling‘s “interaction depth as a scaling axis” claim — MiroThinker reports ~600 tool calls per task at 256K context; Qwen3.7-Max’s announcement claims ~1,000+ at 1M, suggesting the same scaling story holds at frontier scale and closed-weight production deployment.