Qwen3.8-Max Ran Autonomous Chip Design for 500 Turns and Cut 81% of the Circuit
The benchmark numbers from Qwen3.8-Max were reported in the first 24 hours. The technical report detail that followed is the harder-to-summarise part.
The chip design run
Alibaba handed the model a cryptographic chip design task with no reference circuit to copy. Qwen3.8-Max operated for 500 unsupervised turns of edit, simulate, and layout. Its first working circuit used 8,298 logic gates. Over the full 500 turns it squeezed that to 678 gates, a reduction of 91.8% in gate count and 81% in physical chip area, while still meeting 500 MHz timing at the end. No human stepped in between iterations.
The significance: no reference design, no access to prior solutions, no reward signal until functional correctness was confirmed. The model maintained timing constraints as an implicit goal across hundreds of iterations.
The e-commerce simulation
A simulated year of online store operations: 600 suppliers, 7,000 products, 152 fraudulent merchants hidden among them. Qwen3.8-Max started with 100,000 yuan in capital. It ended with 416,252 yuan, roughly 4.2x the starting position. The next best model in the comparison finished 38% lower.
The model had to simultaneously manage supplier relationships, detect and exclude fraudulent actors, and optimise product mix across a year of trading without restarts.
The software development run
Handed an empty folder and a specification for a command-line tool, Qwen3.8-Max ran for 16 days without human intervention. The repository it produced contained 265 commits and 127 pull requests. The model triggered its own build, unit test, and end-to-end validation pipelines after every change — not instructed to do so, but inferring it from the task context.
The paper reproduction task
Given a research paper and access to GPUs, the model reproduced all six of the paper’s experimental findings. It wrote approximately 7,600 lines of code across 5 days and ran 33 rounds of training to get there. The task requires interpreting a dense academic document, writing experiment code from scratch, managing training runs, and verifying results match the reported figures — all unsupervised.
The numbers behind the model
Qwen3.8-Max is a sparse mixture-of-experts architecture with 2.4 trillion total parameters and 95 billion active per token. It prices at $2.00 per million input tokens and $6.00 per million output, with cached reads at $0.17 per million. Context window is 1 million tokens; maximum output is 131,072 tokens with a private thinking budget up to 262,000 tokens.
On standard benchmarks: Terminal-Bench 2.1 at 86.6% (between Claude Opus 4.8 at 84.6% and GPT-5.6 Sol at 88.8%), PaperBench at 93.0%, GPQA Diamond at 92.6%. The model is open-weight.
The pattern
Each of the four tasks shares a structure: long time horizon, sparse feedback, no reference solution, autonomous iteration required. The chip run is 500 turns. The software run is 16 days. The business sim is a full simulated year. These are not single-prompt evaluations.
The question the technical report is implicitly answering is whether a model at this scale can maintain goal coherence across thousands of sequential decisions when corrective feedback arrives only at the end, if at all. For three of the four tasks, the answer appears to be yes.
Key Numbers
| Task | Result |
|---|---|
| Chip design (500 turns) | 8,298 → 678 gates, −81% area, 500 MHz met |
| Business simulation (1 year) | 416,252 yuan from 100,000, +38% vs next best |
| Paper reproduction (5 days) | All 6 findings replicated, 7,600 lines, 33 training rounds |
| Software development (16 days) | 265 commits, 127 PRs, self-initiated CI/CD |
| Terminal-Bench 2.1 | 86.6% |
| PaperBench | 93.0% |
| GPQA Diamond | 92.6% |
| Pricing | $2.00/$6.00 per million tokens |
| Parameters | 2.4T total, 95B active |