Qwen3.5
A model series released by Alibaba's Tongyi Lab under the Qwen family, running in parallel with Qwen3. Detailed parameter specifications have not been fully disclosed, and the series may have connections to the Qwen2.5 line of models.
Core Facts
Qwen (the Tongyi Qwen series) is a family of open-source large models launched by Alibaba.
Qwen3.5's parameter activation rate is 4.3%, while DeepSeek V3's 671B model has an activation rate of 5.5%.
Qwen3.5 uses a hybrid attention architecture that mixes Gated Delta Net (a linear attention variant) with Gated Attention (standard attention).
Timeline (last 90 days)
Qwen3.5 needs targeted constraints for the CPU simulation task; a general-purpose chat model will not spontaneously form a stable CPU state machine.
A project called Qwenputer embeds the behavior of a 32-bit RISC-V CPU into the weights of the Qwen3.5-4B model, enabling the model to execute RISC-V instructions rather than calling an external emulator.
This open-source project has been tested on Qwen3.5 and ToolAlpaca.
The experimental architecture contains three components: a frozen Qwen3.5 backbone, a pointer decision head, and a LoRA adapter.
Qwen3.5 enables thinking mode by default, but the author disables it by default on the server side and lets each caller turn it on as needed, because thinking has no natural stopping point on hard prompts and once burned 35k tokens in a single call.
The commit message states that it implements a decision scorer powered by Qwen3.5, including non-causal attention, publisher readout, image processing, and System One prompt encoding.
PPLX Decider is built on the Qwen3.5 model rather than being a separate small model.
The small model pair halves the download size and reduces peak memory by about 35% compared with the original model pair, but in this trial the full turn was actually slower.
On M4 Max, Omni 3B has a median dialogue generation time of 171 ms and a peak process RSS of 2.57 GiB; Qwen3.5 0.8B has a median dialogue generation time of 96 ms and a peak process RSS of 1.08 GiB; Qwen3-TTS 1.7B has a synthesis time of 1.78 seconds and a peak process RSS of 3.85 GiB.
Qwen3.5, Qwen3.6, and Qwen3.8 gained faster prefill and speculative decoding support on MLX.
28 more timeline events
All Facts (20)
Qwen (the Tongyi Qwen series) is a family of open-source large models launched by Alibaba.
65%VerifiedQwen3.5's parameter activation rate is 4.3%, while DeepSeek V3's 671B model has an activation rate of 5.5%.
65%VerifiedQwen3.5 uses a hybrid attention architecture that mixes Gated Delta Net (a linear attention variant) with Gated Attention (standard attention).
65%UnverifiedQwen3.5 needs targeted constraints for the CPU simulation task; a general-purpose chat model will not spontaneously form a stable CPU state machine.
50%UnverifiedA project called Qwenputer embeds the behavior of a 32-bit RISC-V CPU into the weights of the Qwen3.5-4B model, enabling the model to execute RISC-V instructions rather than calling an external emulator.
50%UnverifiedThis open-source project has been tested on Qwen3.5 and ToolAlpaca.
50%UnverifiedThe experimental architecture contains three components: a frozen Qwen3.5 backbone, a pointer decision head, and a LoRA adapter.
50%UnverifiedQwen3.5 enables thinking mode by default, but the author disables it by default on the server side and lets each caller turn it on as needed, because thinking has no natural stopping point on hard prompts and once burned 35k tokens in a single call.
50%UnverifiedPPLX Decider is built on the Qwen3.5 model rather than being a separate small model.
50%UnverifiedThe commit message states that it implements a decision scorer powered by Qwen3.5, including non-causal attention, publisher readout, image processing, and System One prompt encoding.
50%UnverifiedQwen3.5, Qwen3.6, and Qwen3.8 gained faster prefill and speculative decoding support on MLX.
50%UnverifiedIn evaluations across 3 seeds, 3 models, and 4 benchmarks (involving Gemma 4, Qwen3.5, and Magistral), U-Lens is the best estimator, ahead of 5-round sampling, 10-round sampling, and supervised baselines.
50%UnverifiedNVIDIA Nemotron 3、Qwen3.5 至 Qwen3.8、Kimi-K3 以及 GLM-5.3-Flash 等模型均采用了混合架构范式。
50%UnverifiedQwen3.5 was released under the Apache 2.0 license, permitting commercial use and post-training modifications.
50%UnverifiedQwen3.5 scored 86.7 on TAU2-Bench and 85.0 on MMMU.
50%UnverifiedThe Qwen3.5 flagship 397B-A17B adopts a hybrid architecture that combines Gated Delta Networks with sparse Mixture-of-Experts (MoE), activating only 17 billion parameters per token.
50%UnverifiedThe research team fine-tuned the Qwen3.5 series of small language models on filtered and rebalanced teacher-generated supervised data.
50%UnverifiedBoth the Clef model and the Laya model with a Qwen3.5 backbone can be exported to GGUF format and served via llama.cpp, including multimodal decision models capable of understanding images.
50%UnverifiedThis version supports exporting Clef models based on the Qwen3.5 backbone, as well as Laya models, to GGUF format.
50%UnverifiedQwen3.5 with approximately 800 million parameters achieved the fastest generation speed, approaching 15 tokens per second, with the second-place LFM2.5 exceeding 10 tokens per second.
50%