HauhauCS has released two new uncensored, balanced versions of the Gemma 4 models: Gemma4-26B-A4B and Gemma4-31B-QAT. Both variants incorporate Multi-Token Prediction (MTP) draft heads to enable speculative decoding, resulting in significant inference speed improvements. The 26B-A4B model achieves approximately a 35% speed boost, while the 31B model sees a 53% increase, with identical output quality verified by the model's drafting mechanism. These releases utilize QAT-aware quantization, making Q4_K_M the optimal format as higher precision offers no quality gains for these specific models. The 26B-A4B is a Mixture of Experts architecture with roughly 4 billion active parameters per token, whereas the 31B variant is a dense model offering higher capability for users with sufficient VRAM. Both models include vision support via mmproj files and maintain a 262K context window. The author notes that GenRM testing resulted in zero refusals across 465 prompts, confirming their uncensored nature.
Gemma4-26B-A4B & 31B-QAT Uncensored Balanced Released with MTP Speed Boosts
Specific Labs steers Gemma-4-31B to push back on false premises without benchmark impact
Specific Labs has released Gemma-4-31B-AntiHal, a variant of the Gemma-4-31B model that challenges requests based on false premises rather than hallucinating along with them. The model achieves this through interpretability-based representation steering applied to the residual stream during generation.
Gemma 4 E2B runs at 255 tok/s in browser using WebGPU
Gemma 4 E2B achieves 255 tokens per second in-browser on an M4 Max using WebGPU kernels. The demo and kernels are now available on Hugging Face for public use.
Google releases TimesFM-3, a 330M-parameter model with native multivariate forecasting
Google Research has released TimesFM-3, the third generation of its zero-shot time series foundation model. The primary update from version 2.5 is native support for multivariate inputs, allowing the model to handle multiple simultaneous targets and covariates without fine-tuning.
IFM releases K2-Horizon-MoVA-36B-A4B, a sparse MoE model with 4B active parameters
IFM has released the final checkpoint for K2-Horizon-MoVA-36B-A4B, a Mixture-of-Experts model featuring Mixture-of-Values attention that stores 36B parameters but runs only 4B per token. The release includes GGUF formats and links to smaller variants in the K2-Horizon family.
llama.cpp adds support for NVIDIA Nemotron-3-Puzzle-75B-A9B
llama.cpp introduces support for the NVIDIA Nemotron-3-Puzzle-75B-A9B model, enabling inference for this Mixture of Experts architecture. The update implements infrastructure for variable-per-layer expert feed-forward sizes and top-k routing, which are required because the model has 40 MoE layers with distinct n_ff_exp and top-k values.