Qwen3.5-9B
A 9-billion-parameter model from Alibaba's open-source Qwen series of large language models. It is a mid-scale general-purpose language model suited for scenarios that require a balance between inference cost and model capability.
Timeline (last 90 days)
On the Qwen3.5-4B and 9B models, using only GGA allocation without any other changes improved the scores on seven evaluations by 19.48 and 29.84 percentage points, respectively, compared with STEPQuant@4.
The paper currently validates the seven evaluations only on Qwen3.5-4B/9B, and the method's generalization capability across different model architectures and longer contexts remains to be further observed.
On HealthBench, post-training Qwen3.5-9B with Rubricon rubrics yields performance 4.5% higher than training with RRD, the strongest alternative synthetic rubric generator.
On the LoCoMo benchmark, TrawMem improves the LLM-Judge score over SimpleMem by 4.07 percentage points on Qwen3.5-9B and 4.31 percentage points on DeepSeek-V4-Flash.
Under the condition of matched per-step generation budgets, FOCUS at the Qwen3.5-9B scale improves by 7.2 percentage points over the strongest baseline (the higher of GRPO/VinePPO) on the BrowseComp-Plus benchmark, reaching 55.4 versus 48.2.
The research team conducted fine-tuning experiments on two models, Qwen3.5-9B and Gemma4-12B-it.
Microsoft's Microsoft-Decision-1 claims a decision speed 35 times faster than GPT-6, and its base model is Alibaba's open-source Qwen3.5-9B.
On the Qwen3.5-9B model, SchemaShield-Lite raised accuracy on the DICE-300 benchmark from 0.717 to 0.813, recovering 19 of the 22 drift-induced failures.
Dory scores 61.8 on SWE-bench Verified (4 loops, NVIDIA's own testing), while Qwen3.5-9B scores 61.
CARE is currently validated only on four LIBERO suites and two agents (Qwen3.5-9B and Llama-3.1-8B).
17 more timeline events
All Facts (20)
On the B200, the per-step time for LoRA SFT training of Qwen3.5-9B decreased from 0.84 seconds to 0.77 seconds.
70%UnverifiedThe paper currently validates the seven evaluations only on Qwen3.5-4B/9B, and the method's generalization capability across different model architectures and longer contexts remains to be further observed.
50%UnverifiedOn the Qwen3.5-4B and 9B models, using only GGA allocation without any other changes improved the scores on seven evaluations by 19.48 and 29.84 percentage points, respectively, compared with STEPQuant@4.
50%UnverifiedOn HealthBench, post-training Qwen3.5-9B with Rubricon rubrics yields performance 4.5% higher than training with RRD, the strongest alternative synthetic rubric generator.
50%UnverifiedOn the LoCoMo benchmark, TrawMem improves the LLM-Judge score over SimpleMem by 4.07 percentage points on Qwen3.5-9B and 4.31 percentage points on DeepSeek-V4-Flash.
50%UnverifiedUnder the condition of matched per-step generation budgets, FOCUS at the Qwen3.5-9B scale improves by 7.2 percentage points over the strongest baseline (the higher of GRPO/VinePPO) on the BrowseComp-Plus benchmark, reaching 55.4 versus 48.2.
50%UnverifiedThe research team conducted fine-tuning experiments on two models, Qwen3.5-9B and Gemma4-12B-it.
50%UnverifiedMicrosoft's Microsoft-Decision-1 claims a decision speed 35 times faster than GPT-6, and its base model is Alibaba's open-source Qwen3.5-9B.
50%UnverifiedOn the Qwen3.5-9B model, SchemaShield-Lite raised accuracy on the DICE-300 benchmark from 0.717 to 0.813, recovering 19 of the 22 drift-induced failures.
50%UnverifiedDory scores 61.8 on SWE-bench Verified (4 loops, NVIDIA's own testing), while Qwen3.5-9B scores 61.
50%UnverifiedCARE is currently validated only on four LIBERO suites and two agents (Qwen3.5-9B and Llama-3.1-8B).
50%UnverifiedThe Cloudflare Clef series is entirely open-weight: Clef is based on Qwen3.8-27B, Clef-flash is based on Qwen3.5-9B, and Clef-omni is a 30B mixture-of-experts model (with 3B active parameters).
50%UnverifiedOn LongBench v1/v2 and MemoryAgentBench, NoteBank also reported improvements, covering everything from three 1.3B research models up to Mamba-2-8B and a trained Qwen3.5-9B.
50%UnverifiedSpatialOPSD internalizes privileged depth information into a single RGB forward pass through self-distillation, raising spatial reasoning accuracy from 45.9 to 51.9 on Qwen3.5-9B.
50%UnverifiedReTerm is built on Qwen3.5-9B, and the resulting ReTerm-9B achieves 34.8% Pass@1 on Terminal-Bench 2.0, an improvement of 14.6 percentage points over the base model.
50%UnverifiedIn JetBrains' self-reported evaluations, the 12B MoE Mellum2.1 wins on single-shot code generation tasks (LiveCodeBench, HumanEval+) but loses to the dense 9B Qwen3.5-9B on multi-step agentic benchmarks (SWE-bench Pro, Terminal-Bench).
50%UnverifiedMicrosoft-Decision-1 was built through post-training on Alibaba's Qwen3.5-9B
50%UnverifiedMicrosoft-Decision-1 is priced at $0.042 per million input tokens with no charge for output, is plain-text only, and has a 32k context window.
50%UnverifiedMicrosoft stated that Decision-1 completed its task via post-training based on Qwen3.5-9B and will soon migrate the model to other systems, including Microsoft AI and OpenAI's models.
50%UnverifiedMicrosoft conducted post-training on top of Alibaba's open-source Qwen3.5-9B weights, specifically to perform single-pass decision scoring tasks.
50%