We curate AI news developments for you.
Hugging Face BlogModel release
Hugging Face open-sources AstaBrief 8B, a fast report-generation model for scientific work, which is 3.5× faster than its proprietary counterpart, reducing report generation time to 51.1 seconds. The model and training data are available for researchers to study and adapt.
ReutersModel release
CNBCModel release
Google has been trailing OpenAI and Anthropic for much of 2026, analysts say Gemini 4 puts it back in the frontier AI conversation.
Google: The KeywordModel release
Google released Gemini 4 Argon, a frontier model with a 1-million-token context window for complex tasks like cybersecurity. They also launched new voice models, app integrations, and tools like Googlebook and Gemini for Windows.
CoreWeave BlogModel release
OpenAI's GPT-6 Astra achieved a 57.9% score on Terminal Bench 4.0, outperforming Claude Fable 5.1's 55.8% and costing 63% less per task. Astra also scored 99.9% on ARC-AGI-3 and 0% on ExploitGym.
CoreWeave BlogModel release
Moonshot AI released Kimi K3, a 2.8 trillion-parameter open-weight model that overwhelmed compute capacity and sparked controversy over its origins. The model features a 1 million token context, 896 routed experts, and benchmarks top in coding and automation tasks at lower costs.
Mistral AI NewsModel release
Mistral AI released Leanstral 1.5, a free Apache-2.0 model with 6B active parameters, achieving 587/672 on PutnamBench and 87% on FATE-H. It uncovered 5 bugs in 57 repositories and is available via Hugging Face and a free API.
Mistral AI NewsModel release
Mistral AI released Robostral Navigate, an 8B model that enables robot navigation with a single RGB camera, achieving 76.6% success on R2R-CE benchmarks, outperforming multi-sensor approaches. It uses pointing-based navigation and reinforcement learning, operating on wheeled, legged, and flying robots.
Cohere BlogModel release
Cohere launched Transcribe, an open-source speech recognition model with 5.42% word error rate, ranking #1 on HuggingFace's Open ASR Leaderboard. It supports 14 languages and is available via Hugging Face and Cohere's Model Vault.
Cohere BlogModel release
Cohere released Command A+, an open-source language model optimized for agentic tasks, running on as few as two H100 GPUs. It outperforms previous models with 218B total parameters and 25B active parameters, achieving 63% higher output tokens per second.
Mistral AI NewsModel release
Mistral AI released Shieldstral, a 3B open-weights multimodal safety classifier that outperforms models up to 7x its size. It runs on a single 16GB GPU and evaluates text and images using plain-language policies at inference time.
Cohere BlogModel release
Cohere launched North Mini Code, a 30B-parameter agentic coding model with 3B active parameters, open-source under Apache 2.0. It achieved a 33.4 score on the Artificial Analysis Coding Index and 2.8x higher throughput than Devstral Small 2.
Cohere BlogModel release
Cohere released North Small Translate, a machine translation model that achieved an 83.6 WMT26 score, outperforming models like DeepL and Google Translate. It offers high throughput and cost efficiency, with a $0.000676 per task cost.
Mistral AI NewsModel release
Mistral AI launched Voxtral TTS, a 4B-parameter text-to-speech model supporting 9 languages with realistic, emotionally expressive speech. It offers low latency, voice adaptation in 3s, and starts at $0.016 per 1k characters.
Mistral AI NewsModel release
Mistral AI released Mistral Small 4, a unified model combining reasoning, coding, and multimodal capabilities with 119B parameters and a 256k context window. It offers 40% lower latency and 3x higher throughput than its predecessor.
Cohere BlogModel release
Cohere released Embed 5, a new family of embedding models with Pro and Fast tiers, offering improved retrieval performance for enterprise data. Embed 5 Pro achieved an 85.8 score on ViDoRe V3, while the Fast tier costs $0.08 per million tokens.
Mistral AI NewsModel release
Mistral AI released OCR 4, an OCR model with bounding boxes, block classification, and confidence scores, supporting 170 languages and running in a single container. It outperforms leading systems in human evaluations and is priced at $4 per 1,000 pages via API.
Meta AI BlogModel release
Meta launched Muse Image and previewed Muse Video, advanced media generation models. Muse Image, ranked No. 2 on Arena for image tasks, supports precise editing and multi-reference composition, and is available on Meta AI, Instagram Stories, and WhatsApp.
Meta AI BlogModel release
Meta introduced Muse Spark 1.1, a multimodal reasoning model with improved tool use, coding, and multimodal understanding, available via the Meta Model API. It features a 1 million token context window and enhanced agentic capabilities for complex tasks.
The InformationModel release
Microsoft on Thursday launched speech-generating AI that the company says is both cheaper and more accurate than competing models from the likes of ElevenLabs, SpaceXAI and Google. Microsoft aims to produce voice models to power features such as transcribing conference calls in its Teams app or ...
NVIDIA BlogModel release
NVIDIA Blackwell GPUs power OpenAI's GPT-6 Astra Ultrafast, offering up to 8x faster token generation than standard mode. The feature is available via the OpenAI API for eligible users.
Microsoft SourceModel release
Microsoft launched MAI-Transcribe-2-Streaming, ranking No. 1 in accuracy, and two voice models: MAI-Voice-2.1 and MAI-Voice-2.1-Flash, supporting 23 languages with native accents. MAI-Transcribe-2-Streaming transcribes in 60 languages at 2x faster speed than competitors, priced at $0.54 per hour.
CNBCModel release
Google is promising major advances in coding and cybersecurity, but the company is quickly falling behind in personal agents.
CNBCModel release
Gemini 4 Argon is Alphabet's most advanced model yet, with major coding, cybersecurity, and complex professional work improvements.
TechCrunchModel release
Amazon Web Services released Strands Decider 2B, an open-source decision model inspired by Jev, built on Qwen3.5-2B. It offers fast, low-cost decision-making with confidence scores.
Cloudflare BlogModel release
Cloudflare released Clef and Clef-flash, open-source decision models, which outperformed Jev in benchmarks and offer faster processing. Clef processed a domain in 2.2 seconds, compared to 4.7 seconds for a general LLM.
Cloudflare BlogModel release
Cloudflare announced two European open models, EuroLLM and Apertus, now available on Workers AI. EuroLLM supports 35 languages, and Apertus was trained on 1,500 languages with 15 trillion tokens.
TechCrunchModel release
Google released Gemini 4 Argon, claiming it is its most powerful model yet, designed for cybersecurity, coding, and research, and rolled out to select partners via the Fairwind Program. It outperformed OpenAI's GPT-6 Astra and Anthropic's Fable and Opus models in AI benchmarks, according to Google.
The VergeModel release
Google launched Gemini 4 Argon, limiting access to "trusted cyber defenders" while ensuring it is not misaligned. The model outperforms competitors in benchmarks and is already used internally for tasks like codebase migrations.
Google DeepMind BlogModel release
Google launched Gemini 4 Argon, a model with a 1 million token limit for complex tasks, rolling out to trusted cyber defenders via the Fairwind Program. It excels in coding, finance, legal work, and cybersecurity, with a $2 per million input token price.