Qwen3.5-4B is the 4-billion-parameter model in the Qwen3.5 family. Recent coverage has focused on derivatives built on it: Microsoft released FrogNano-4B-2609, post-trained with reinforcement learning on repository-level software engineering tasks, while a consultant's LoRA fine-tune ImaJev-4B was reported first of 91 on JevBench. Other projects embed a 32-bit RISC-V CPU inside its weights and offer a neuron explorer.
Developer embeds 32-bit RISC-V CPU inside Qwen model weightsA developer built a 32-bit RISC-V CPU directly inside the weights of the Qwen3.5-4B model, letting it execute instructions internally. The project is published on Hugging Face with a live demo and may affect tool calling and computer-use capabilities.
Open-source Neuron Explorer released for Qwen3.5-4BA neuron explorer for Qwen3.5-4B was released on GitHub, pushing every MLP neuron up and down on prompts and laying results out in a 3D model view. It supports live steered chat, lens analysis and connection inspection locally.
Developer previews 4B coding model that generates 3D scenesA developer shared an early 4B model fine-tuned for general coding and 3D scene generation, an independent tune of Qwen3.5-4B. Q4_K_M and Q8_0 GGUF versions are on Hugging Face. Training is ongoing; the model and dataset will be updated and released soon.
Microsoft releases FrogNano-4B agentic modelMicrosoft released FrogNano-4B-2609 on Hugging Face, a model derived from Qwen3.5-4B and post-trained on repository-level software engineering with reinforcement learning across about 1,500 synthetic SWE task environments. Aimed at users with limited GPU resources, its performance depends on the Leaf harness and test quality, and generated patches need human review.
Developer Fine-Tunes 4B Model to Top Decision-Making BenchmarksA business consultant spent 15 days fine-tuning a Qwen3.5-4B model with LoRA and a small decision head to create ImaJev-4B, which handles text and up to two images and returns probabilities per option. It ranked #1 of 91 on JevBench (67.37) and #3 of 56 on DecisionBench, ahead of GPT-5.6 Luna and DeepSeek V4.1. Total cost about $1,200; weights released under Apache-2.0.
Open-source Kev 4B tested side by side with JevA side-by-side test of Kev 4B (an Apache-2.0 Qwen3.5-4B fine-tune) and Jev on 362 fresh items found accuracy within 2 points on every task. Jev led on PAWS paraphrase detection (87.0% vs 74.5%) but counts ~257 extra input tokens per request, making short requests up to 12x costlier.
Four AI Models Tested on Doom Control BenchmarkAn experiment let Jev, Laya, finetuned ModernCE and finetuned Qwen3.5-4B play ViZDoom autonomously, comparing kills, survival time and latency. Finetuned Qwen3.5-4B averaged 3.63 kills, the best, while Jev survived 5.63s on average with p95 latency near 200ms. No model received Doom-specific training.
The A.I. Daily