Qwen3.5 is a large language model in Alibaba's Qwen family, widely used by developers as a base for fine-tuning and local inference. Recent coverage centers on local and small-model experiments: a project compiled an NES emulator into roughly 1GB of Qwen3.5 weights, Nvidia built its open-source Flora data engine on three fine-tuned Qwen3.5 models, and Bilibili released the Qwen3.5-based Index-Translate family. Other stories cover running Qwen3.5-9B INT4 on cheap FPGA mining cards, small decision models such as Kev and Jeff, and GGUF support in transformers targeting the Qwen3.5 architecture.
The nesformer project has compiled the NES 6502 CPU, its work RAM and cartridge ROM into the weights of a stock Qwen3.5 model, with every value computed by a compiler rather than trained. The model predicts the CPU's execution log token by token while a host program handles graphics and input; test games match a conventional 6502 emulator frame for frame. It builds on Qwenputer, reusing its compiler and Gated DeltaNet state as machine memory.
Nvidia has open-sourced its data engine Flora, which labels an hour of robot video for just $0.72 versus $88 with GPT-6 Astra, while achieving slightly higher quality. Built on three fine-tuned Qwen3.5 models, Flora more reliably catches errors in robot actions and is a key piece of Nvidia's strategy to control the robotics value chain.
The open-source openTPU accelerator, developed by AI agents, spans SystemVerilog hardware, an instruction set, a bit-exact simulator, a kernel language and compiler, and host software, running ten real-weight models on a Xilinx Kintex-7 FPGA card with bit-exact token output versus the simulator.
A developer used ~$280 SQRL FK33 ex-mining FPGAs, with Claude assisting the RTL work, to run Qwen3.5-9B INT4 inference. Two cards at 75MHz produced ~3.2 tok/s generation, dropping to ~2.4 tok/s at longer context, with layer-by-layer validation against llama.cpp.
Bilibili open-sourced Index-Translate, a translation model family built on Qwen3.5. Text models cover 150 languages and extend to speech translation, syllable control and full-document translation.
A NeurIPS 2026 paper finds LLMs that push back on wrong users still accept the same wrong answer from a 'verified source,' an effect called Authority Bias. One source note flipped 45-88% of correct answers in 7 of 8 models; GPT-5.4 flipped 44.7% and Grok-4.20 87.5%.
The independent open project Jeff releases 0.8B and 2B zero-shot decision models fine-tuned from Qwen3.5 and Gemma 4, returning calibrated probabilities in a single forward pass at about 22 ms on an RTX PRO 6000 and 28 ms on an M4 Max, trained only on local hardware with synthetic data and released on Hugging Face.
Hugging Face added GGUF support to transformers, letting users load Hub checkpoints sized for their machine via from_pretrained. To approach llama.cpp performance, it reuses ggml kernels and cuts generate overhead, initially targeting Apple Silicon and the Qwen3.5 architecture. Local tools like Ollama, LM Studio and Jan are powered by llama.cpp.
Developer jaredpalmer released Kev, a family of small decision models built on Qwen3.5, offered in 0.8B, 4B and 9B parameter sizes with pretrained weights, training code and evaluation data. The models handle yes/no, multiple-choice and rating questions, run on CUDA and Apple Silicon (4B and 9B fit a 32GB Mac in bf16), and expose an API matching TypeSafe System One.
Cactus Compute open-sourced Needle 3, a 121M-parameter automation foundation model that runs offline on-device. Trained on 360B tokens, it uses a Simple Attention / Hadamard MLP design with 70.8M parameters stored as engrams, reducing compute to ~100 MFLOPs/token. On Mobile Actions it scores 86.0, outperforming much larger models, and is available on Hugging Face, GitHub, PyPI, plus a browser sandbox.
That is everything
The A.I. Daily
Project compiles NES emulator into ~1GB of Qwen weights
