TechnologyTracking
Does Microsoft's Decision-1 (Qwen3.5-9B, JevBench 83.5%) show Chinese base models are internationally competitive?
Microsoft's Decision-1 model scored 83.5% on JevBench but is post-trained on Alibaba's Qwen3.5-9B—does this indicate that Chinese base models are now internationally competitive for decision-making tasks?
Current read
Microsoft Decision-1, post-trained on Alibaba's Qwen3.5-9B, leads in 36 tests including our assessmentBench, indicating Chinese base models' competitiveness in decision tasks, though benchmark authority and independent verification remain unconfirmed.
Evidence confirms Microsoft released Decision-1 in October 2026, post-trained on Alibaba's Qwen3.5-9B for structured decision tasks, outputting calibrated probabilities rather than text. Microsoft claims it topped 36 benchmarks including our assessmentBench, is 35x faster than GPT-6, and passed 11 safety tests; it is available on Microsoft Foundry and OpenRouter. These facts indicate Chinese base models' competitiveness in decision tasks, but our assessmentBench's authority, independent verification, and real-world generalization remain unspecified, so the conclusion should be cautiously bounded.
How we read it
Microsoft Decision-1, post-trained on Qwen3.5-9B, topped 36 benchmarks, indicating Chinese base models' competitiveness in decision tasks.
The model is available on Microsoft Foundry and OpenRouter, but our assessmentBench's authority and independent verification are not specified in the reports.
Earlier updates
- — Public tracking started; no independent probability estimate yet.; Starting public context: Microsoft released Microsoft-Decision-1, a decision model post-trained on Alibaba's Qwen3.5-9B, on October 12. Microsoft claims it scored 83.5% across 36 our assessmentBench benchmarks, surpassing Quyet-1.0-Large's 81.9%, with 85-millisecond median latency and 4.5x speed over the runner-up. The model is available on OpenRouter and Microsoft Foundry, priced at $0.042 per million input tokens with free output. Microsoft says it is testing the model internally for incident response, quality control and scientific discovery.
- — Microsoft Decision-1, post-trained on Alibaba's Qwen3.5-9B, leads in 36 tests including our assessmentBench, indicating Chinese base models' competitiveness in decision tasks, though benchmark authority and independent verification remain unconfirmed.; Evidence confirms Microsoft released Decision-1 in October 2026, post-trained on Alibaba's Qwen3.5-9B for structured decision tasks, outputting calibrated probabilities rather than text. Microsoft claims it topped 36 benchmarks including our assessmentBench, is 35x faster than GPT-6, and passed 11 safety tests; it is available on Microsoft Foundry and OpenRouter. These facts indicate Chinese base models' competitiveness in decision tasks, but our assessmentBench's authority, independent verification, and real-world generalization remain unspecified, so the conclusion should be cautiously bounded.