MIT researchers found that OpenAI’s GPT-6 Astra can inspect and directly rewrite neural network weights without any gradient descent or retraining. In one test, Astra modified just 0.016% of a small Qwen model’s weights to eliminate repetitive loops, bumping its math score from 7.9% to 48.1%. If frontier models can reliably program weight space, AI self-improvement might arrive much faster than expected.
ScenemaAI just open-sourced Scenema Audio, an expressive speech model extracted from the LTX 2.3 audiovisual transformer. It lets you direct emotional shifts and ambient background noise using stage-direction action tags, alongside zero-shot voice cloning. The model fits in 16GB VRAM and supports 13 languages.
Google just introduced TranslateGemma, an open translation model family in 4B, 12B, and 27B sizes built on Gemma 3. Distilled from Gemini, the 12B model outperforms the earlier 27B baseline across 55 languages while keeping multimodal support to translate text inside images.
Google DeepMind just open-sourced Gemma Scope 2, an interpretability suite designed to reverse-engineer how the Gemma 3 family actually thinks. Spanning models from 270M to 27B parameters, it uses sparse autoencoders and transcoders across every layer to map internal computations. Why it matters: It gives the safety community an open, microscopic look into why models hallucinate, get jailbroken, or disguise their reasoning.
Google just launched T5Gemma 2, an encoder-decoder model family built on Gemma 3 that introduces multimodal vision and massive 128K token contexts. By tying embeddings and merging attention mechanisms, Google squeezed parameters down to compact sizes (~370M to ~7B) ideal for on-device deployment. It is available as pre-trained weights ready for downstream fine-tuning.
Google just released FunctionGemma, a lightweight model tuned from Gemma 3 270M specifically for on-device function calling. Small enough to run locally on phones and Jetson Nano boards, fine-tuning boosts its task accuracy from 58% to 85%. Why it matters: It turns edge devices into fast, fully private agents capable of triggering native actions without touching the cloud.
Google just launched EmbeddingGemma, a compact 308M parameter open embedding model tailored specifically for on-device AI. It allows developers to run semantic search and local RAG pipelines entirely offline, keeping sensitive data private while sharing a tokenizer with Gemma 3n to save memory. High-accuracy retrieval is finally fitting inside lightweight local hardware.
Google just released Gemma 3 270M, an ultra-compact 270-million-parameter open model built specifically for task-specific fine-tuning. Packed with out-of-the-box instruction following and structuring capabilities, it lets developers ditch bloated models in favor of cheap, fast specialized workers for jobs like text classification and data extraction.
Google just released MedGemma 27B Multimodal and a 400M-parameter image encoder called MedSigLIP for health AI development. Both run on a single GPU—with lightweight variants even fitting on mobile devices—to help researchers build private, tailored medical tools without relying on closed APIs.
Google just released optimized versions of its Gemma 3 open models using Quantization-Aware Training (QAT). By simulating low-precision operations during training, the 4-bit (int4) weights cut memory demands drastically while slashing perplexity loss by 54%. This means developers can now run powerful 27B models locally on consumer cards like the NVIDIA RTX 3090 instead of expensive cloud clusters.
Google just released Gemma 3, its newest lineup of open-weight models ranging from 1B to 27B parameters. The 4B, 12B, and 27B versions handle both images and text with an expanded 128k token context window, supporting over 140 languages. Why it matters: The 27B model beats Gemini 1.5 Pro across key benchmarks, offering top-tier multimodal power you can run openly.
