Gemini 3.5 Flash-Lite
Gemini 3.5 Flash-Lite is a low-latency, lower-cost Gemini model, made generally available on July 21, 2026, for subagent tasks, document parsing and high-volume workflows. It takes text, image, video, audio and PDF input and returns text.
Released
July 21, 2026
Type
multimodal
License
proprietary
Context
1,048,576 tokens
Pricing
Input
$0.30 / 1M tokens
Output
$2.50 / 1M tokens
List price on the developer's own pages, checked . Paid tier, Standard. Input price is the same for text, image, video and audio. Output includes thinking tokens. Context caching $0.03. Free tier available.
Capabilities
Sources
Every figure on this page was checked against these pages on . A figure we could not source is left out.
- Gemini API docs: Gemini 3.5 Flash-Lite (opens in a new tab)
- Gemini API changelog (opens in a new tab)
- Gemini API deprecations (opens in a new tab)
- Gemini API pricing (opens in a new tab)
Spotted something out of date? Tell us.
Links
In the news: Gemini 3.5 Flash-Lite and Google DeepMind
MarkTechPost
Google Releases Gemini 4 Argon API With 1M Token Output Limit
Google DeepMind has made its Gemini 4 Argon API generally available to developers, enterprises, and paid Gemini subscribers. Gemini 4 Argon is a sparse mixture-of-experts frontier model built for long, multi-step work, with target domains including coding, finance, legal, science, and cybersecurity. The model was first announced on September 30 for a select group of cyber defenders and is now accessible via the paid Gemini API at introductory pricing of $2 per input million tokens and $10 per output million tokens. A key enhancement for developers is the increased output length, with Argon capable of generating up to 1 million tokens in a single response, a significant increase from the 64K limit on earlier Gemini models. The model card indicates a combined input and output window of up to 1 million tokens, meaning a lengthy prompt will reduce available output space. Gemini 4 Argon is described as a new, large-scale sparse MoE transformer with native multimodal support, accepting text, images, audio, and video inputs, and producing text outputs. It was trained on TPUs using JAX and ML Pathways, with a knowledge cutoff of March 2026. Google has not disclosed parameter counts. Performance benchmarks show Argon leading Google's internal tables for knowledge work, long context, and video understanding, achieving 91.7% on LVBench. However, it performed less strongly on agentic coding tasks, scoring 55.0% on FrontierSWE v2, placing it last among four models in Google's comparison.
Hugging Face
One Model Family, Two Gold-Level Results: Fine-Tuning Nemotron for IOI and IMO
Google DeepMind's Nemotron model family has achieved gold medal-level performance in two prestigious international competitions: the International Olympiad in Informatics (IOI) and the International Mathematical Olympiad (IMO). This achievement marks a significant milestone for AI in competitive problem-solving, demonstrating advanced capabilities in algorithmic reasoning and mathematical problem-solving. The Nemotron models, specifically fine-tuned for these challenging tasks, were able to tackle complex problems that typically require human-level intellect and strategic thinking. The IOI, which focuses on computer science and algorithmic problem-solving, saw Nemotron excel in areas such as data structures, algorithm design, and computational complexity. The IMO, on the other hand, tested Nemotron's prowess in abstract reasoning, number theory, geometry, and combinatorics. The specific details of the fine-tuning process and the exact Nemotron model variants used were not disclosed in the initial announcement, but the success indicates a sophisticated approach to adapting large language models for specialized domains. This dual success highlights the growing potential of AI systems to not only assist humans but also to compete and excel in fields previously considered exclusive to human cognitive abilities. The development of Nemotron by Google DeepMind, a leading AI research laboratory known for its work on models like AlphaFold and AlphaGo, underscores the company's commitment to pushing the boundaries of artificial intelligence. The ability of these models to achieve top-tier results in academic competitions suggests a deeper understanding and application of logic and reasoning than previously demonstrated by AI. The implications of this development are far-reaching. In education, AI models like Nemotron could serve as powerful learning tools, helping students understand complex concepts and develop problem-solving skills. In research, they could accelerate discovery by assisting scientists and mathematicians in exploring new theories and solutions. However, it also raises important questions about the future of human competition, the definition of intelligence, and the ethical considerations surrounding AI's increasing capabilities in academic and professional arenas. Further analysis of the specific problem sets and Nemotron's solutions will be crucial to fully understand the extent of this AI achievement.
MarkTechPost
Google DeepMind Releases EmbeddingGemma 2, a 740M Open Multimodal Embedding Model Built on Gemma 4
Google DeepMind has released EmbeddingGemma 2, an open-source multimodal embedding model designed to convert text, code, images, video, and audio into a unified 768-dimensional vector space. This model, built upon the Gemma 4 architecture, features 740 million parameters and boasts an 8K token context window, representing a significant upgrade from its predecessor. The release is accompanied by an Apache 2.0 license, making it freely available for use and modification. EmbeddingGemma 2 is engineered for applications such as on-device search, classification tasks, and privacy-first Retrieval Augmented Generation (RAG) pipelines. The model's ability to generate embeddings locally ensures that sensitive data remains on the user's device, thereby reducing latency and enabling offline functionality. The core innovation of EmbeddingGemma 2 lies in its capacity to create a single vector space that encompasses multiple modalities. This means a text query can retrieve relevant images, or a voice memo can be used to find a corresponding video clip. The model is modular, comprising a text and code backbone with 270 million parameters (further divided into a 130M transformer and a 140M embedder), a 170M parameter vision encoder, and an optional 300M parameter audio encoder. Developers can choose to load only the components they require, leading to a flexible parameter count: 270M for text-only, 440M for text and vision, 570M for text and audio, or the full 740M for all modalities. Crucially, all configurations share a single vector space, allowing embeddings generated by a text-only setup to match documents embedded by the complete model. The context window has been expanded to 8,192 tokens, which is four times larger than that of EmbeddingGemma 1. This increased capacity can accommodate approximately 29 images, 58 video frames, or 5.5 minutes of audio, enhancing its ability to process richer and longer inputs. Google's research team has reported that EmbeddingGemma 2 achieves leading scores among multimodal embedders with fewer than 1 billion parameters on the MTEB Code and MAEB benchmarks. Specifically, it scored 78.68 on MTEB Code v1, an improvement from EmbeddingGemma 1's score of 68.76, and 61.36 on MTEB multilingual v2, slightly surpassing EmbeddingGemma 1's 61.15. The model's weights are now accessible on platforms like Hugging Face and Kaggle, with builds available for Ollama, llama.cpp GGUF, and LiteRT, facilitating immediate deployment by developers.
MarkTechPost
Mistral AI Releases Mistral Large 4 (Le Chonk): A 1.05T Parameter Multimodal MoE Model
Mistral AI announced the public preview release of Mistral Large 4 (ML4), internally codenamed 'Le Chonk,' on October 6, 2026. This new model is a granular Mixture of Experts (MoE) architecture, boasting a total of 1.05 trillion parameters. Specifically, it features 49 billion active parameters per token, a 1.6 billion parameter vision encoder for native image input, and an extensive 1 million token context window. The model was trained from scratch using 3,800 NVIDIA Grace Blackwell GPUs within Mistral's own European data centers. The API for ML4 is now accessible, with pricing set at $1.36 per 1 million input tokens and $4.18 per 1 million output tokens. However, the model weights are scheduled for release at the end of October, meaning self-hosting capabilities will not be available until then. Mistral AI highlights ML4's standout performance in cybersecurity, reporting a 93% score on the Cybench benchmark and 82% on CyberGym-E2E. The company noted that several closed frontier models achieved near-zero scores on these benchmarks due to task refusal. The architecture of ML4 is described as a hybrid instruct-and-reasoning MoE that natively accepts image inputs. The efficient activation of only approximately 4.7% of the total weights per token allows a model of this scale to be served at mid-tier pricing. While the full 1.05 trillion parameters require memory, the activation count dictates the compute resources needed, not the overall hardware expenditure for users. Mistral has not yet disclosed the specific expert count, top-k routing mechanisms, or layer layout, details that are expected to be provided upon the release of the model weights. The training data for ML4 encompassed over 160 languages, including all official European Union languages, indicating a broad linguistic capability. Mistral AI, founded in 2023 by former Google DeepMind and Meta AI researchers, is a European artificial intelligence company focused on developing open and efficient large language models. The company has previously released models like Mistral 7B and Mixtral 8x7B, which gained traction for their performance and accessibility. The development of ML4 signifies a significant step forward in the company's ambition to compete with larger, established AI labs by pushing the boundaries of model scale and multimodal capabilities while maintaining a focus on efficiency and European data sovereignty. The release of ML4 positions Mistral AI as a key player in the competitive landscape of advanced AI development, particularly with its emphasis on specialized benchmarks like cybersecurity.
MIT Technology Review
People really hate AI, so why can’t they get enough?
A significant portion of the global population expresses apprehension and nervousness regarding artificial intelligence, yet simultaneously demonstrates rapidly increasing engagement with AI products and services. This dichotomy suggests a complex and evolving public relationship with AI technology. A survey by the Pew Research Center indicates that more adults in the United States anticipate a negative personal and societal impact from AI than a positive one, with pessimism being particularly pronounced among younger demographics. Further underscoring this sentiment, a report from Stanford University revealed that over half of people worldwide feel nervous about AI products and services. Public opposition to AI infrastructure is also notable; a Gallup poll conducted in May found that 71% of US adults would oppose the construction of a new AI data center in their vicinity, a figure exceeding opposition to new nuclear power plants by a considerable margin. An NBC poll from March even placed AI's popularity below that of internal combustion engine (ICE) vehicles. Despite these widespread concerns, the adoption and usage of AI tools are experiencing exponential growth. OpenAI's ChatGPT reached a milestone of one billion monthly users in May, according to market analysis firm Sensor Tower. Google DeepMind's AI model, Gemini, is closely following, having recorded 950 million users in July. The Pew Research Center's data further illustrates this trend, showing that half of US adults now report using a chatbot, more than double the number from 2023, with one in four users engaging with these tools daily. This pattern is not confined to the United States; globally, over one-third of adults across all 38 OECD countries have utilized generative AI tools within the past three months. This widespread adoption, even amidst expressed anxieties, highlights the perceived utility and perhaps the addictive nature of these emerging technologies. The CEO of Springboards, a startup developing a large language model (LLM) designed for a broader range of responses than mainstream competitors, articulated a sentiment of "self-loathing" within the AI industry, suggesting an internal awareness of the technology's potentially problematic aspects. This internal reflection mirrors the external public sentiment, where a "love/hate" attitude towards AI has become prevalent. The technology is described as being "warped by hype, steered by zealots, inescapable," indicating a perception of AI's development and deployment being driven by factors beyond pure utility or societal benefit. The contrast between public unease and escalating usage points to a societal grappling with the profound implications of AI, balancing perceived risks against undeniable conveniences and capabilities. The data from various research institutions and market analysis firms consistently paints a picture of a public simultaneously wary and captivated by the advancements in artificial intelligence, creating a unique and challenging landscape for developers, policymakers, and users alike.