{"company":{"name":"DeepMind","slug":"deepmind","website":"https://deepmind.com","category":"ai-ml"},"question":"What has DeepMind shipped recently?","answer":"In the last 30 days, DeepMind shipped 24 tracked updates. The most recent was \"Image understanding¶\" on 2026-11-17.","window":{"days":30,"updateCount":24,"returned":24},"generatedAt":"2026-11-18T00:37:06.911Z","updates":[{"title":"Image understanding¶","summary":"This Gemma developer guide explains how to use Gemma 3 and later models for image understanding, including describing image content, identifying objects, recognizing scenes, and inferring context. It walks through loading the model with the Hugging Face transformers library, prompting with single and multiple images, optical character recognition for multilingual text, and object detection with bounding boxes on a normalized 1000x1000 grid that users must rescale to the original image size. The guide also describes variable resolution in Gemma 4, where a user-selected token budget of 70, 140, 280, 560, or 1120 visual tokens trades inference speed against visual accuracy. A worked example shows that a 70-token budget misses some cars and people in object detection. The page is a how-to tutorial rather than an announcement, so it reports no new product, pricing, or policy change.","date":"2026-11-17","dateIsEstimated":true,"signalType":null,"signalTypeLabel":null,"sourceUrl":"https://gemma.ai/get-started/capabilities/image/","publisher":"gemma.ai"},{"title":"Audio understanding¶","summary":"DeepMind's Gemma documentation describes audio input support in Gemma 3n and Gemma 4, covering automatic speech recognition, speech translation, and general speech understanding via Hugging Face Transformers. The guide states that Gemma 4 E2B, E4B, and 12B Unified are trained for multilingual transcription and translation. It also gives audio preprocessing requirements, such as converting input to mono 16 kHz float32 waveforms, and shows a microphone capture example in a notebook.","date":"2026-11-17","dateIsEstimated":true,"signalType":"feature_update","signalTypeLabel":"Feature","sourceUrl":"https://gemma.ai/get-started/capabilities/audio/","publisher":"gemma.ai"},{"title":"Speed-up Gemma 4 with Multi-Token Prediction¶","summary":"Gemma 4 introduces a Multi-Token Prediction (MTP) drafter, a smaller model that shares embeddings and last-layer activations with the target model to propose several tokens for parallel verification. The page states this yields significant decoding speedups with output quality identical to standard autoregressive generation, aimed at low-latency and on-device use. It notes the MoE variant (Gemma 4 26B A4B) may not speed up at batch size 1 on hardware with poor parallelism. Developers can load the drafter as an assistant model in Hugging Face transformers and tune the number of draft tokens or use a heuristic schedule.","date":"2026-11-17","dateIsEstimated":true,"signalType":null,"signalTypeLabel":null,"sourceUrl":"https://gemma.ai/get-started/capabilities/mtp/","publisher":"gemma.ai"},{"title":"Diffusion in Text Generation Explained¶","summary":"Gemma.ai published a technical explainer on DiffusionGemma, a text-generation model that applies diffusion techniques to language rather than producing one token at a time. The page argues that standard autoregressive Transformers are memory-bound during decoding, so a single user gets no latency benefit from idle compute. DiffusionGemma instead initializes a 256-token canvas of random tokens and refines the whole canvas in parallel, shifting the workload to compute-bound. It uses Uniform State Diffusion, which replaces words with random vocabulary tokens and re-noises low-confidence tokens so errors can be corrected in later steps. The model is a fine-tuned Gemma 4 26B A4B backbone that toggles between incremental prefill and denoising modes. Self-conditioning carries prediction memory between steps, and multi-canvas block sampling chains 256-token diffusion blocks autoregressively for long-form output. The page is educational and includes no launch date, pricing, or availability details.","date":"2026-11-17","dateIsEstimated":true,"signalType":null,"signalTypeLabel":null,"sourceUrl":"https://gemma.ai/models/diffusiongemma/explained/","publisher":"gemma.ai"},{"title":"Generate output","summary":"DiffusionGemma is described as an experimental open model exploring text diffusion, built on a 26B (4B active) Mixture-of-Experts Gemma 4 architecture and generating tokens via discrete diffusion. The guide covers loading the model with the transformers library, prompt templates, system instructions, thinking mode, and image understanding, and states that running it requires a GPU with more than 60GB of memory such as an NVIDIA H100.","date":"2026-11-17","dateIsEstimated":true,"signalType":null,"signalTypeLabel":null,"sourceUrl":"https://gemma.ai/models/diffusiongemma/inference/","publisher":"gemma.ai"},{"title":"Function calling with Hugging Face Transformers¶","summary":"A technical tutorial explains how to use FunctionGemma, described as a specialized version of the Gemma 3 270M model trained for improved function calling, within the Hugging Face ecosystem. It covers installing torch and transformers, loading the model with AutoProcessor and AutoModelForCausalLM, and passing tools either as manual JSON schemas or raw Python functions. The guide advises using manual schemas for complex custom objects, since automatic conversion may describe nested objects only as generic \"object\" types. It also notes that a developer system message is needed so the model generates structured calls. The page does not state a release date.","date":"2026-11-17","dateIsEstimated":true,"signalType":"content_marketing","signalTypeLabel":"Content","sourceUrl":"https://gemma.ai/models/functiongemma/inference/","publisher":"gemma.ai"},{"title":"Gemini Omni 1.2 Flash","summary":"Your browser does not support the video tag. Your browser does not support the video tag. Gemini Omni 1.2 Flash Greater creative control and production-ready outputs Learn more Try","date":"2026-11-17","dateIsEstimated":true,"signalType":"feature_update","signalTypeLabel":"Feature","sourceUrl":"https://aistudio.google.com/prompts/new_chat?model=gemini-omni-1.2-flash&utm_source=deepmind.google&utm_medium=referral&utm_campaign=gdm&utm_content=","publisher":"aistudio.google.com"},{"title":"Google Cloud K8s Engine adds Gemma deployment guides for vLLM, TensorRT-LLM, and TPU JetStream","summary":"Google Cloud Kubernetes Engine now offers deployment guides for running DeepMind's Gemma models using vLLM, TensorRT-LLM on GPUs, and JetStream on TPUs, along with data analysis and fine-tuning pipelines. These integrations enable high-performance, low-latency serving and customization of Gemma models.","date":"2026-11-13","dateIsEstimated":true,"signalType":"integration","signalTypeLabel":"Integration","sourceUrl":"https://ai.google.dev/gemma/docs/core/gke","publisher":"ai.google.dev"},{"title":"Introducing Gemini 4 Hydrogen","summary":"Introducing Gemini 4 Hydrogen","date":"2026-11-12","dateIsEstimated":false,"signalType":null,"signalTypeLabel":null,"sourceUrl":"https://deepmind.google/blog/introducing-gemini-4-hydrogen/","publisher":"deepmind.google"},{"title":"Introducing Gemini 4 Hydrogen","summary":"DeepMind introduced Gemini 4 Hydrogen, a new model optimized for rapid execution in agentic workflows and everyday coding, surpassing Argon on the Artificial Analysis Intelligence Index. It is priced at $1.50 per million input tokens and $7.50 per million output tokens across a 1M-token context window and is available globally in Google AI Studio, the Gemini API, and consumer products like the Gemini app and Google Search.","date":"2026-11-12","dateIsEstimated":false,"signalType":"product_launch","signalTypeLabel":"Launch","sourceUrl":"https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-hydrogen/?utm_source=deepmind.google&utm_medium=referral&utm_campaign=gdm&utm_content=","publisher":"blog.google"},{"title":"AlphaGenome 2: Deciphering how cells read DNA to better understand disease","summary":"DeepMind introduced AlphaGenome 2, an advanced AI model designed to predict how DNA sequences and their variations influence human cell types and disease. This update enhances the ability to decode genetic mechanisms, potentially accelerating disease research and therapeutic development.","date":"2026-11-11","dateIsEstimated":false,"signalType":"feature_update","signalTypeLabel":"Feature","sourceUrl":"https://deepmind.google/blog/alphagenome-2-deciphering-how-cells-read-dna-to-better-understand-disease/","publisher":"deepmind.google"},{"title":"Gemini 4 Carbon is now available","summary":"Gemini 4 Carbon is now available","date":"2026-11-10","dateIsEstimated":false,"signalType":null,"signalTypeLabel":null,"sourceUrl":"https://deepmind.google/blog/gemini-4-carbon-is-now-available/","publisher":"deepmind.google"},{"title":"Gemini 4 Carbon is now available","summary":"DeepMind introduced Gemini 4 Carbon, its strongest frontier model, now available to developers, enterprises, scientists, and consumers. Carbon excels in cybersecurity, long-horizon agentic workflows, and complex scientific reasoning, outperforming competitors in benchmarks while offering lower costs. It integrates with Google Antigravity for multi-agent collaboration and is priced at $2 per million input tokens and $10 per million output tokens until November 30, 2026.","date":"2026-11-10","dateIsEstimated":false,"signalType":null,"signalTypeLabel":null,"sourceUrl":"https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-carbon-now-available/?utm_source=deepmind.google&utm_medium=referral&utm_campaign=gdm&utm_content=","publisher":"blog.google"},{"title":"Opening the datasets behind our AI weather models","summary":"Foundational precipitation and geostationary satellite datasets to observe the planet clearly.","date":"2026-11-10","dateIsEstimated":false,"signalType":"content_marketing","signalTypeLabel":"Content","sourceUrl":"https://deepmind.google/blog/opening-the-datasets-behind-our-ai-weather-models/","publisher":"deepmind.google"},{"title":"Gemini Robotics 2.5 accelerates real-world general dexterity","summary":"DeepMind introduced Gemini Robotics 2.5, its most capable vision-language-action model, designed to enhance robot dexterity, reliability, and safety in real-world applications. The model accelerates progress toward general physical intelligence for robots.","date":"2026-11-09","dateIsEstimated":false,"signalType":"product_launch","signalTypeLabel":"Launch","sourceUrl":"https://deepmind.google/blog/gemini-robotics-2-5-accelerates-real-world-physical-intelligence/","publisher":"deepmind.google"},{"title":"Gemini Omni 1.2 Flash: Fast and cost-effective with longer clips","summary":"DeepMind introduced Gemini Omni 1.2 Flash, a faster and more cost-effective version of its model that supports longer video clips. The update enhances processing efficiency while reducing operational costs, making it more accessible for developers and enterprises. Existing users can adopt the new model immediately.","date":"2026-11-05","dateIsEstimated":false,"signalType":"feature_update","signalTypeLabel":"Feature","sourceUrl":"https://deepmind.google/blog/gemini-omni-12-flash-fast-and-cost-effective-with-longer-clips/","publisher":"deepmind.google"},{"title":"Agentic cheminformatics with Google Antigravity","summary":"DeepMind introduced agentic cheminformatics capabilities in Google Antigravity, enabling accelerated drug discovery workflows. The update allows searching compounds and proposing novel rigidified macrocycles, enhancing AI-driven research in chemistry.","date":"2026-11-03","dateIsEstimated":false,"signalType":"feature_update","signalTypeLabel":"Feature","sourceUrl":"https://deepmind.google/blog/agentic-cheminformatics-with-google-antigravity/","publisher":"deepmind.google"},{"title":"Models - Gemini API | Google AI Studio","summary":"DeepMind introduced Gemini 4 Argon as its new flagship reasoning model with a shared 1M-token context window and up to 1M-token output, alongside a broad expansion of the Gemini 3 lineup including new Flash TTS, Live API, and image-generation models. The update also deprecates Gemini 2.5 Flash with a migration deadline of December 1, 2026.","date":"2026-11-02","dateIsEstimated":true,"signalType":null,"signalTypeLabel":null,"sourceUrl":"https://aistudio.google.com/docs/models","publisher":"aistudio.google.com"},{"title":"Speeding up planning decisions with AI","summary":"DeepMind and the UK government announced a partnership to deploy AI tools that reduce planning assessment times for infrastructure and development projects. The collaboration aims to streamline decision-making processes, though specific tools or timelines were not detailed in the announcement.","date":"2026-10-30","dateIsEstimated":false,"signalType":"partnership","signalTypeLabel":"Partnership","sourceUrl":"https://deepmind.google/blog/speeding-up-planning-decisions-with-ai/","publisher":"deepmind.google"},{"title":"What is work for? Human wellbeing and the value of work in the AGI era","summary":"DeepMind published research examining the non-monetary value people derive from work, including purpose, social status, and belonging, across four countries. Findings show many would not leave jobs even for 120% of their salary, highlighting the inadequacy of income transfers alone to address AI-driven labor disruption.","date":"2026-10-29","dateIsEstimated":true,"signalType":"content_marketing","signalTypeLabel":"Content","sourceUrl":"https://institute.deepmind.com/essays/what-is-work-for-human-wellbeing-and-the-value-of-work-in-the-agi-era/","publisher":"institute.deepmind.com"},{"title":"Gemini 4 Carbon: Our next frontier model","summary":"Gemini 4 Carbon: Our next frontier model","date":"2026-10-28","dateIsEstimated":false,"signalType":null,"signalTypeLabel":null,"sourceUrl":"https://deepmind.google/blog/gemini-4-carbon-our-next-frontier-model/","publisher":"deepmind.google"},{"title":"Gemini 4 Carbon: our next frontier model","summary":"DeepMind introduced Gemini 4 Carbon, its strongest model to date, excelling in cybersecurity and long-horizon tasks. It leads benchmarks like SEC-bench Pro (88.2%) and CWE-bench (71.0%), and shows a 2x uplift in secure code generation over Argon. Carbon will roll out gradually via the Fairwind Program before broader availability at $2/$10 per million tokens (introductory pricing).","date":"2026-10-28","dateIsEstimated":false,"signalType":null,"signalTypeLabel":null,"sourceUrl":"https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/?utm_source=deepmind.google&utm_medium=referral&utm_campaign=gdm&utm_content=","publisher":"blog.google"},{"title":"Genie 4: A new generation of world simulation","summary":"DeepMind introduced Genie 4, a next-generation world model designed to deliver unprecedented realism in real-time simulations. The model builds on prior generations to enhance dynamic environment generation and interaction fidelity.","date":"2026-10-27","dateIsEstimated":false,"signalType":null,"signalTypeLabel":null,"sourceUrl":"https://deepmind.google/blog/genie-4-a-new-generation-of-world-simulation/","publisher":"deepmind.google"},{"title":"EmbeddingGemma 2: Unified and Efficient Multimodal Representations","summary":"DeepMind introduced EmbeddingGemma 2, an open multimodal embedding model that unifies text, images, audio, video, and combinations in a single representation space. The model outperforms larger specialist models in embedding benchmarks while offering flexibility for resource-constrained environments through selective encoder loading and quantized weights.","date":"2026-10-27","dateIsEstimated":false,"signalType":null,"signalTypeLabel":null,"sourceUrl":"https://deepmind.google/research/publications/279443/","publisher":"deepmind.google"}],"attribution":{"source":"Spyingbee","url":"https://spyingbee.com/updates/deepmind","citation":"Spyingbee, \"DeepMind updates\", https://spyingbee.com/updates/deepmind (retrieved 2026-11-18)"}}