Skip to content
IFAI

AI models

New releases, benchmarks, pricing and API changes from GPT, Claude, Gemini, DeepSeek, Qwen and other large language models, updated daily.

11 stories in the last 7 days

  1. Reach 53

    OpenAI and Meta team up to teach AI African languages

    OpenAI and Meta are partnering with telecom giant Orange to fine-tune AI models for regional African languages currently ignored by mainstream GenAI. The project starts with West Africa's Wolof and Pulaar, training Whisper and Llama 3.1 to power voice support, with plans to release models under a free non-commercial license for local health and education. Why it matters: It’s a practical push for digital inclusion, bringing AI to millions who don't speak dominant global languages.

  2. Reach 66

    Nvidia's Fugatto creates sounds you've never heard

    Nvidia just debuted Fugatto, a generative audio model that transforms existing sounds and invents entirely new soundscapes from text prompts. It can morph piano chords into singing voices or turn a train rumble into an orchestra. Nvidia is withholding its public release for now over safety and copyright concerns.

  3. Reach 45

    Samsung drops Gauss2 AI model in three distinct sizes

    Samsung just revealed Gauss2, its next-generation AI model family spanning three tiers: Compact for on-device computing, Balanced for speed and efficiency, and Supreme for heavy workloads. The models support up to 14 spoken languages and multiple coding languages, and already power internal tools like code.i. Samsung plans to weave Gauss2 into its wider consumer device lineup next.

  4. Reach 60

    Google's Gemini-exp-1121 beats GPT-4o on coding and vision

    Google just dropped Gemini-exp-1121, an experimental multimodal model that outperforms OpenAI's GPT-4o by roughly 20% across coding, math, and vision benchmarks. It features deeper contextual reasoning and optimized architectures, making it much sharper at turning raw design sketches into functioning code.

  5. Reach 57

    OpenAI beefs up GPT-4o with better writing and file analysis

    OpenAI just rolled out an update to GPT-4o, significantly improving its creative writing chops and file handling performance. Uploaded PDFs and text docs now get deeper, more detailed analysis, while generated text and marketing copy sound noticeably more natural and human-like.

  6. Reach 53

    DeepSeek takes on o1 with new reasoning preview

    DeepSeek just launched DeepSeek-R1-Lite-Preview, delivering OpenAI o1-preview-level performance on AIME and MATH benchmarks. The model lets users watch its thought process unfold in real time, with open-source weights and API access slated to arrive soon.

  7. Reach 46

    Niantic is using Pokémon Go player scans to train spatial AI models

    Niantic just announced it is building a Large Geospatial Model trained on millions of real-world 3D scans uploaded by Pokémon Go and Ingress players. The goal is to give robots and computers human-like spatial intelligence, using pedestrian-level maps of places cars cannot reach.

  8. Reach 45

    How Zuckerberg rebuilt Meta around Llama

    Mark Zuckerberg just bet Meta's future on open-source AI, turning the Llama model series into the backbone of its apps and smart glasses. After the metaverse slump, downloads topped 600 million, though investors are still waiting to see how billions in computing spend will pay off.

  9. Reach 54

    Stable Diffusion 3.5 Large lands on AWS SageMaker

    Stable Diffusion 3.5 Large is now available on Amazon SageMaker JumpStart, bringing the 8-billion-parameter text-to-image model directly to enterprise cloud setups. It generates 1-megapixel images across multiple styles and deploys securely inside your VPC to protect proprietary data.

  10. Reach 42

    Microsoft rolls out industry-tailored AI models built on Phi SLMs

    Microsoft just launched a suite of industry-adapted AI models developed with partners like Bayer, Siemens, and Cerence. Built on its lightweight Phi architecture, these specialized SLMs target practical domain tasks from crop protection and in-car controls to factory floor troubleshooting via Azure AI Studio.

  11. Reach 60

    The AI scaling wall is getting real

    AI giants OpenAI, Google, and Anthropic are reportedly hitting a wall while developing next-generation models. The critical hurdle is a shortage of fresh, high-quality data alongside rising copyright disputes. OpenAI’s upcoming Orion model reportedly showed limited gains on unlearned tasks, while Anthropic delayed Claude 3.5 Opus. The brute-force scaling era may finally be running out of internet text.

  12. Reach 51

    OpenAI shifts strategy as Orion AI progress slows

    OpenAI is rethinking its development playbook as its next flagship model, code-named Orion, shows smaller performance gains than anticipated. Early tests indicate it struggles to deliver dramatic improvements in areas like coding. The company is now leaning into synthetic data and post-training methods, while confirming Orion will not arrive this year.

  13. Reach 54

    Sam Altman predicts AGI will arrive in 2025

    OpenAI CEO Sam Altman just bet on 2025 as the year we reach artificial general intelligence. In a YC interview, he claimed the path to AGI is now basically clear and boils down to engineering rather than scientific breakthroughs, even as reports swirl about LLM scaling challenges.

  14. Reach 51

    Google brings Gemini to the OpenAI SDK

    Google just added native compatibility for Gemini models directly inside the OpenAI Library and REST API. Developers can now switch to Gemini using Chat Completions and Embeddings endpoints without rewriting code—simply point the base URL to Google and plug in a Gemini API key.

  15. Reach 50

    OpenAI’s new trick to make GPT-4o edits faster

    OpenAI just launched Predicted Outputs, a new API capability that slashes latency for gpt-4o and gpt-4o-mini by letting you feed in a reference string. It is tailored for partial rewrites—like updating docs, iterating on prior chat answers, or refactoring code inside an existing file.

  16. Reach 51

    Anthropic quadruples the price of its new Haiku model

    Anthropic just launched Claude 3.5 Haiku, but it comes with a surprise 4x price hike and no image analysis at launch. While the new model matches or beats its previous flagship Claude 3 Opus on key benchmarks, input tokens now cost four times more than before. Anthropic will keep Claude 3 Haiku around for cost-conscious users.

  17. Reach 57

    Meta opens Llama AI models to U.S. military

    Meta just made an exception to its military ban, allowing U.S. defense and national security agencies to use its Llama models. Partners like Lockheed Martin and Oracle are already tapping the AI for code generation and aircraft maintenance. Meta is also extending access to government partners in the U.K., Canada, Australia, and New Zealand.

  18. Reach 51

    OpenAI leaders drop hints on autonomous agents and AGI in Reddit AMA

    OpenAI executives just hosted a Reddit AMA, revealing that autonomous agents capable of reaching out first and taking action on users' behalf will be a "big theme in 2025." CEO Sam Altman also claimed AGI is achievable with current hardware, while teasing upcoming upgrades to ChatGPT Search, DALL-E, and HIPAA compliance for enterprise customers.

  19. Reach 61

    Claude can now see the pictures in your PDFs

    Anthropic just updated Claude 3.5 Sonnet to analyze PDFs visually in public beta, meaning it can now read text alongside charts, tables, and embedded images. Files can be up to 100 pages and 32 MB, with processing available via Claude Chat preview and the API. It makes digesting complex financial reports and slide decks much easier.

  20. Reach 51

    Gemini API gets live Google Search grounding

    Google just added Grounding with Google Search to Google AI Studio and the Gemini API, letting Gemini 1.5 fetch live web answers with inline citations. Testing is free inside AI Studio, while API access runs $35 per 1,000 grounded queries. Why it matters: developers can finally plug Google's web index directly into Gemini without messy custom retrieval pipelines.