Qwen3.5-9B
6 storiesQwen3.5-9B is a 9-billion-parameter open model in Alibaba's Qwen family, and recent coverage shows it being used as a base for other work. Microsoft built its Decision-1 scoring model on it for routing, classification and verification, while developers fine-tuned it into decision models Bev and Jeeves. The garlic-llm project also runs it offline on a stock PS4 using custom 4- and 3-bit repacked weights at roughly 31–39 tok/s.
Related topics
The open-source garlic-llm project runs Qwen3.5-9B offline on a stock PS4 GPU at 31–39 tok/s. Speculative decoding with the model's built-in MTP head lifted speed from 22 tok/s, and weights were repacked to custom 4- and 3-bit formats to fit about 4.5 GiB.
Microsoft released Decision-1, built on Alibaba's Qwen3.5-9B, designed to score fixed answer options with probabilities rather than generate open-ended text. Available via Microsoft Foundry and OpenRouter, it targets routing, classification, and verification.
Microsoft introduced Microsoft-Decision-1, a decision-scoring model for routing, classification, prioritization and verification, available in Microsoft Foundry and soon on OpenRouter. Microsoft says it topped a 36-benchmark comparison and was fastest, 4.5x quicker than runner-up Quyet-1.0-Large and 35x than GPT-6 Sol; it was post-trained from Qwen3.5-9B.
A developer fine-tuned Qwen3.5-9B into Bev, a decision model that answered correctly only 1.9% of 324 held-out decisions while averaging 96% confidence. Built as a control case, it tests whether automated pipelines actually verify model correctness rather than just confidence.
A new paper introduces Context Language Models (CLMs), which treat context as a file the model manages on its own. Zero-shot CLMs beat state-of-the-art context-management strategies: 11.4% higher accuracy with 21.5% fewer FLOPs on BrowseComp-Plus, plus gains on EdgeBench and a 24-hour multi-repository task. An online RL method and Suffix Cache Reuse serving further cut compute.
A developer released Jeeves, a Jev-like decision model built on Qwen3.5-9B with LoRA, a pointer head, CISPO training, and a block-4 diffusion drafter. It scores 0.889 on unseen test data versus 0.822 for Kev-9B and 0.857 for Jev, and 0.935 on JevBench's public hard tier; latency on one H100 is about 0.3s without thinking and 3.3s median with it, with code, weights, and data open-sourced.
That is everything