Qwen3.7 Max
Qwen3.7 Max by Alibaba is a text-only reasoning agent model with 1M-token context, cross-harness training, preserve_thinking API, and closed-weight API-only access.
Model Overview
Capabilities, design details, and architectural traits
Qwen3.7 Max - Alibaba's Text-Only Agent Frontier Model
Qwen3.7 Max is officially framed by Alibaba under the title "The Agent Frontier". It is a text-only reasoning model - it accepts text input and returns text output only, with no vision capability. Its training methodology explicitly separates the task, the agentic harness, and the verifier as distinct components to prevent the model from learning tricks tied to any single agent framework.
Documented Traits That Distinguish Qwen3.7 Max
| Trait | Detail |
|---|---|
| Cross-harness agent training | Trained on many combinations of task, harness, and verifier - explicitly designed to prevent single-harness overfitting and deliver stable performance across diverse agent frameworks |
preserve_thinking API parameter | Retains <think> reasoning blocks across conversation turns, preventing loss of reasoning chain between tool calls in long-horizon sessions |
| 1M-token context window | Text-only; 64K output token limit; context window doubled from Qwen3.6 Max Preview's 256K |
| Native OpenAI and Anthropic API compatibility | Documented at launch; drops into existing OpenAI-compatible or Anthropic-compatible integrations without infrastructure changes |
| Closed weights | API-only access via Alibaba Cloud Model Studio - a documented break from the prior Qwen open-weight release pattern |
How Cross-Harness Training Defines the 3.7 Generation
Alibaba's documented approach separates three typically coupled components in agent training: the task, the agentic harness (the tool-calling scaffold), and the verifier (the success evaluator). Training across many combinations of these components prevents the model from learning patterns specific to one setup. This is the official explanation for Qwen3.7 Max's documented ability to generalize across frameworks including OpenClaw, Hermes Agent, and Qwen Code without per-framework tuning.
Benchmark Performance
Independent evaluations · Artificial Analysis
Accuracy & Capability Details
Independent evaluation data provided by Artificial Analysis. To view the latest benchmarks and full details, visit their official site.
Compare Models Side-by-Side
Evaluate specifications, pricing, and independent benchmark indices
| Model Details | AlibabaHost Qwen3.7 Max | ||
|---|---|---|---|
| General Info | |||
| Provider | Alibaba | Anthropic | Anthropic |
| Release Date | May 19, 2026 | June 9, 2026 | May 28, 2026 |
| Knowledge Cutoff | — | — | — |
| Context & Limits | |||
| Context Window | 1M | 1M | 1M |
| Pricing (per 1M tokens) | |||
| Input Pricing | $2.50 Best Input Pricing | $10 | $5 |
| Output Pricing | $7.50 Best Output Pricing | $50 | $25 |
| Modalities | |||
| Inputs | text | textimagefile | textimagefile |
| Outputs | text | text | text |
| Benchmarks (0-100) | |||
| Intelligence Index | 46.0 | 59.9 Best Intelligence Index | 55.7 |
| Coding Index | 66.0 | 76.5 Best Coding Index | 74.3 |
| Agentic Index | 30.6 | 52.8 Best Agentic Index | 47.2 |
GPQA Benchmark
Graduate-level reasoning and expert Q&A evaluation.
Humanity's Last Exam
Extremely difficult logical reasoning and knowledge.
IFBench Instruction
Evaluation of strict instruction following.
Benchmark scores are independent evaluations sourced from Artificial Analysis. Intelligence, Coding, and Agentic indices reflect composite performance ratings (0–100).