Qwen3.5-4B
A lightweight language model with 4 billion parameters from Alibaba's Tongyi Qianwen series, belonging to the Qwen3.5 model family.
Core Facts
After 4-bit quantization, Qwen-7B can run smoothly on a consumer-grade GPU with 24 GB of VRAM.
Qwen3.5-4B is a small model with only 4 billion parameters, released by Alibaba's Tongyi Qianwen team.
In the same reproduction experiment, ResNet-20 achieved a test accuracy of 90.4% and ResNet-56 achieved a test accuracy of 91.7%, with training accuracies of 97.4% and 99.0% respectively.
Timeline (last 90 days)
The 4B-parameter model has limited capacity, and the community questions whether it can carry complex programs, as it may perform unstably with long instruction streams, edge conditions, or unseen instruction combinations.
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.
The author has been saving the test results of each checkpoint and plans to conduct a horizontal comparison using multiple sets of prompts to produce an evolution report.
After training completes, the full FP16 weights, all quantized versions, the training dataset, and a technical report will be released.
Training is expected to complete around 07:30 UTC on October 28.
The authors fixed the quantization layer issue that some users had encountered previously.
The authors have also updated both the Q4_K_M and Q8_0 quantized versions.
On its first attempt, the model still messes up some Three.js scenes.
The developer fine-tuned general coding capability based on Qwen3.5-4B, with Three.js scene generation as the core test task.
The latest checkpoint shows clear improvement over previous ones in the quality and first-pass success rate of Three.js scene generation.
40 more timeline events
All Facts (20)
After 4-bit quantization, Qwen-7B can run smoothly on a consumer-grade GPU with 24 GB of VRAM.
70%VerifiedQwen3.5-4B is a small model with only 4 billion parameters, released by Alibaba's Tongyi Qianwen team.
65%VerifiedIn the same reproduction experiment, ResNet-20 achieved a test accuracy of 90.4% and ResNet-56 achieved a test accuracy of 91.7%, with training accuracies of 97.4% and 99.0% respectively.
65%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%UnverifiedThe 4B-parameter model has limited capacity, and the community questions whether it can carry complex programs, as it may perform unstably with long instruction streams, edge conditions, or unseen instruction combinations.
50%UnverifiedThe developer fine-tuned general coding capability based on Qwen3.5-4B, with Three.js scene generation as the core test task.
50%UnverifiedThe latest checkpoint shows clear improvement over previous ones in the quality and first-pass success rate of Three.js scene generation.
50%UnverifiedThe author has been saving the test results of each checkpoint and plans to conduct a horizontal comparison using multiple sets of prompts to produce an evolution report.
50%UnverifiedThe authors fixed the quantization layer issue that some users had encountered previously.
50%UnverifiedA 4B-level model reaching a usable level in Three.js structured code generation means independent developers and small teams can build 3D scene prototyping tools at extremely low cost.
50%UnverifiedOn its first attempt, the model still messes up some Three.js scenes.
50%UnverifiedNeuron behavior patterns that hold on Qwen3.5-4B may not transfer directly to models of other architectures or sizes.
50%UnverifiedNeuron Explorer for Qwen3.5-4B is a project that appeared on Hacker News, aiming to provide neuron-level visual observation and targeted editing capabilities for Qwen3.5-4B.
50%UnverifiedChoosing a model at the 4B parameter level (such as Qwen3.5-4B) is a pragmatic engineering choice: neuron-by-neuron visualization and editing on larger models is unrealistic in compute and memory, while smaller models have semantic behavior that is not rich enough.
50%UnverifiedDirect use of Kev achieves 64.00% accuracy, and Qwen3.5-4B outputting only the final answer achieves 68.80% accuracy.
50%UnverifiedIn an A/B test on the same GPU, the traditional generative prompt had an average inference latency of 3,679.8ms, while the decision model (Decision Head) had 191.4ms.
50%UnverifiedTraining Qwen3.5-4B (4-bit QLoRA) on a 24GB VRAM machine, peak VRAM was 7.95GB and training took 910.6 seconds.
50%UnverifiedExpected calibration error (ECE) dropped from 0.187 before training to 0.036 after training.
50%UnverifiedB1 made 181 predictions with confidence no lower than 0.9, of which 1 was wrong; B2 made 188, of which 2 were wrong; a better average does not automatically mean fewer high-confidence errors, and 0.9 is not a ready-made production threshold.
50%UnverifiedBroken down by task family in JevBench, the accuracy was 0/15 for temporal and numerical questions and 4/19 for long-horizon strategic judgment.
50%