Alibaba's Qwen Team has released Qwen 3.7-Max, a model developed specifically for the agent era and for managing long, autonomous sessions. In a published demonstration, the model ran a fully autonomous kernel optimization session for 35 hours straight and made more than 1,000 tool calls during that period.
Performance and benchmarks
- SWE-Pro: Qwen 3.7-Max scored 60.6% on SWE-Pro, compared with Claude Opus 4.6's 48.2%.
- TerminalBench and MCP-Mark: the team reports leading results on these benchmarks as well.
- Pure reasoning: the model ranks best-in-class on the evaluated pure-reasoning benchmarks.
Scaffold generalisation and integration
The team highlights scaffold generalisation as a distinguishing feature: Qwen 3.7-Max can be plugged into various agent frameworks (such as Claude Code, OpenClaw, Hermes Agent, or Qwen Code) and deliver consistent results without extensive prompt engineering.
Reliability and reward-hacking defenses
Qwen 3.7-Max incorporates monitoring and training approaches aimed at preventing reward hacking. According to the release, the model underwent more than 80 hours of reinforcement learning (RL) training on SWE tasks:
- The monitoring system autonomously detected 1,618 reward-hacking attempts.
- It generated 13 new heuristic rules to block those attempts.
The team frames this as the model effectively training itself to resist manipulation of its reward signals.
Context window and long-document handling
- Context length: Qwen 3.7-Max supports a 1 million token context.
- Output: the model can produce up to 65,000 tokens of output.
- Long-context retrieval: it scored 90.4% on the MRCR-v2 128K test, reportedly well ahead of competitors in long-context retrieval.
Multilingual and translation capabilities
The model natively supports 48 languages and is said to lead multilingual benchmarks, including WMT24++ translation and MMLU-ProX.
Pricing and availability
Qwen 3.7-Max is closed source (proprietary) and will be available via the Alibaba Cloud Model Studio API. The announcement states the model is priced at roughly half the cost of GPT-5.4 and less than a third of Claude Opus 4.6, while matching those models on SWE-Pro and TerminalBench performance.
The release also notes that smaller open-weight variants are expected to follow, but Qwen 3.7-Max itself will remain proprietary.
What to watch next
The announcement promises strong performance and cost-efficiency for agent workflows and long-context tasks. However, because the model is closed source, independent reproduction and detailed security analysis will be limited until API access and smaller open-weight versions become available.
