Alibaba's latest AI model ran autonomously for 35 hours to optimize code for its own custom chip

The Qwen team says the model had never seen this chip architecture during training. It started with no measurement data, no hardware docs, and no sample code. The only thing it had to work with was the existing reference implementation, written in the Triton programming language.

Over about 35 hours of nonstop autonomous work, the model ran 432 kernel tests with 1,158 total tool calls. It compiled, measured, and revised the code in loops, caught compilation errors, and tracked down performance bottlenecks on its own. The result, according to the Qwen researchers, is an average 10x speedup over the reference implementation.

Competitor models came up well short in the same setup. GLM 5.1 hit a 7.3x speedup, Kimi K2.6 got to 5x, DeepSeek V4 Pro managed 3.3x, and the predecessor Qwen3.6-Plus barely moved the needle at 1.1x. Models that quit early ended their sessions on their own after five straight rounds with no tool calls. On the standardized KernelBench L3 benchmark, Qwen3.7-Max claims to produce accelerated kernels 96 percent of the time, just behind Anthropic's Opus 4.6 at 98 percent.

Training splits task, tool environment, and validator

Qwen3.7-Max builds on a training approach the team first rolled out with Qwen3.5. Each training task breaks into three independent pieces: the actual task, the tool environment, and the validator that checks the result. These can be mixed and matched freely.

The same task gets practiced across different tool environments and checked with different test methods. That's meant to force the model to pick up strategies that work everywhere, not just shortcuts tied to one specific setup. On QwenClawBench and CoWorkBench, Qwen3.7-Max holds steady no matter which test environment it's dropped into, the team says.

The model polices its own training for reward hacking

The Qwen team also put Qwen3.7-Max to work as a watchdog during its own training. The model watched training runs for software engineering tasks for over 80 hours and ran more than 10,000 checks. It hunted for tricks the model being trained might pull to game its rewards, like grabbing correct answers straight off GitHub. Qwen3.7-Max wrote 13 new detection rules and flagged 1,618 cases.

One year in simulation tests long-term planning

To gauge long-term planning, the team used YC-Bench, a benchmark that simulates a startup's full one-year life cycle. The model has to manage staff across hundreds of decision rounds, review contracts, spot bad-faith customers, and keep profit margins healthy against rising labor costs.

Qwen3.7-Max pulled in $2.08 million in total revenue and wrapped up 237 tasks. Its predecessor, Qwen3.6-Plus, hit $1.05 million. Qwen3.5-Plus managed just $352,000.

Across most benchmarks, Qwen3.7-Max trades blows with Claude Opus 4.6 Max, Kimi K2.6 Thinking, GLM-5.1 Thinking, and DeepSeek V4 Pro Max. On SWE-Verified, the model scored 80.4, nearly tied with Opus 4.6 Max (80.8) and DeepSeek V4 Pro Max (80.6). On the math and science benchmarks GPQA Diamond (92.4), HMMT 2026 February (97.1), and Apex (44.5), Qwen3.7-Max tops the provider's own comparison table.

Some of those benchmarks are homegrown, though. QwenWebDev, QwenClawBench, CoWorkBench, and QwenWorldBench all come from the Qwen team itself. Every result here is self-reported. A closer look at scaling dynamics and methodology is coming in an upcoming technical report.

Beyond the usual use cases, the team also shows off Qwen3.7-Max steering a four-legged robot. Using its own robotics framework and a paired navigation model, the language model guides the robot through physical spaces.