Qwen 3.5 Medium
Qwen 3.5 Medium is a family of open-weight language models released by Alibaba’s Tongyi Lab on February 24, 2026. The series comprises four variants: the mixture-of-experts Qwen3.5-122B-A10B (122 billion total parameters, 10 billion active) and Qwen3.5-35B-A3B (35 billion total, 3 billion active), the dense Qwen3.5-27B, and the lightweight Qwen3.5-Flash. All models support a 128,000-token context window and are distributed under the Apache 2.0 license. In published benchmarks, the 122B-A10B configuration achieves 74.2 on MMLU-Pro and 81.3 on HumanEval, compared to 71.8 and 78.5 for the prior Qwen 2.5-235B-A22B (235 billion total, 22 billion active). On MATH-500, it scores 90.2, and on LiveCodeBench it reaches 48.9, exceeding the 22B-active predecessor by 2.8 and 5.1 points respectively. This release is significant because it demonstrates that a 10-billion-active-parameter open model can match or surpass a predecessor with over twice the active parameters across reasoning, mathematics, and code generation, while its permissive license enables self-hosted deployments that directly challenge proprietary mid-tier models.
Alibaba’s Tongyi Lab dropped Qwen 3.5 Medium on February 24, 2026, a family of open-weight MoE models (122B/10B active and 35B/3B active) that directly challenge Meta’s Llama lineage and Nemotron-Cascade 2. The model has been absorbed fast: Kiro and Qwen-Scope already use it, and Amazon’s SageMaker now supports agentic fine-tuning for it. A May 16 vLLM optimization slashed voice AI latency by 40% on a 6-GPU cluster, signaling production readiness. Alibaba also opened its Qwen app to external partners via a China Eastern deal. Sparse autoencoders on the 27B variant exposed 81k features—a transparency play. But mention counts are flat (zero in last 7/30 days), and the model competes on the same open-weight turf as Meta. The question: can Qwen 3.5 Medium sustain momentum when the next Llama drops?
- ·Released Feb 24, 2026 by Alibaba's Tongyi Lab as open-weight MoE family
- ·Competes directly with Meta and Nemotron-Cascade 2
- ·Adopted by Kiro and Qwen-Scope products; Amazon SageMaker supports it
- ·May 16 vLLM optimizations reduced voice AI latency by 40% on 6 GPUs
- ·Zero mentions in last 30 days despite strong initial deployment velocity
Signal Radar
Five-axis snapshot of this entity's footprint
Mentions × Lab Attention
Weekly mentions (solid) and average article relevance (dotted)
Timeline
5- Executive ChangeMar 5, 2026
Leadership exodus at Qwen AI team with technical lead and multiple staff members leaving
View source - Research MilestoneFeb 25, 2026
Outperformed its 235B parameter predecessor while using 7x fewer active parameters per token
View source- parameter count:
- 35B
- efficiency gain:
- 7x fewer active parameters
- Research MilestoneFeb 24, 2026
Demonstrated remarkable efficiency gains through architectural improvements
View source - Product LaunchFeb 1, 2026
Recently released model used for performance comparison
Relationships
6Competes With
Frequently appears with
10Entities that show up in the same articles — shared coverage, not a stated relationship.
Recent Articles
No articles found for this entity.
Predictions
No predictions linked to this entity.
AI Discoveries
1- observationactiveJul 12, 2026
Lifecycle: Qwen 3.5 Medium
Qwen 3.5 Medium is in 'declining' phase (0 mentions/3d, 0/14d, 24 total)
90% confidence