194 posts on this topic
Anthropic has officially introduced Claude Fable 5.2 and Claude Mythos 5.2, delivering 30% faster token generation, a 20% API price cut, and breakthrough capabilities in coding and science.
Microsoft has filed a patent for a distributed computing platform that pools idle graphics cards from personal computers to train artificial intelligence models, rewarding users with cash, store credits, or subscriptions.
The countdown has begun for Qwen 4, Alibaba Cloud's next-generation AI model family. The stealth Qoduck model inside Qoder IDE, leaks in the vLLM codebase and the ultra-sparse MoE architecture tested with Qwen3.8-Flash-Next reveal what the Qwen 4 family will offer.
Speaking on the Cheeky Pint show, OpenAI CEO Sam Altman pointed to the global memory bottleneck and argued that RAM sitting idle in personal computers for most of the day should be pooled in data centers instead, sparking a major debate in the tech world.
AI giant Anthropic has formally launched its initial public offering (IPO) with an S-1 registration statement filed with the US Securities and Exchange Commission (SEC). The financials of the company, which is preparing to trade on Nasdaq under the ticker ANT, reveal the enormous scale of the generative AI economy.
Apple is rapidly expanding the language options for Siri AI, its new-generation AI-powered digital assistant. Having rolled out five new languages with the iOS 27.2 update, the tech giant announced that Turkish language support will also be officially released by the end of the year.
Moonshot AI has unveiled Kimi K3.1 Pro, its flagship open-weight multimodal reasoning model. Built on a Sparse Mixture-of-Experts (MoE) architecture with 2.8 trillion total parameters (104 billion active per token) and a 1 million token context window, the model performs on par with closed-source rivals in long-horizon software engineering, parametric 3D/CAD modeling and financial analysis tasks while offering 33% lower API costs.
NVIDIA has unveiled Nemotron 3.5 Bolt E4B, a new open-weight model built for edge devices, robotic systems and personal computers. Built on Google's Gemma 4 E4B architecture and running with 8 billion total and 4 billion active parameters, the model offers high-speed inference fully offline, support for 35 languages and multi-turn tool-calling capabilities with a requirement of just 4 GB of VRAM.
DeepSeek V4.1-Flash, which sits at the top of open-weight AI models, now completes tasks much faster while spending fewer tokens with its new update, and competes head to head with closed-source flagship models.
DeepSeek has released DeepInfer, the high-performance inference framework it built to run massive production workloads, to the open-source world under the Apache 2.0 license. We take a close look at the architecture of the engine, which supports NVIDIA SM100 GPUs and Huawei Ascend 950 NPUs, its DSpark speculative decoding capabilities, and the missing pieces developers need to build to set up a full live serving stack.
Chinese tech giant Huawei and AI lab DeepSeek are running a broad strategic partnership, from chip design to the open source TileLang infrastructure, to strengthen the domestic hardware and software ecosystem against US sanctions.
Chinese AI company DeepSeek has released V4.1-Flash-1017, an MIT-licensed open-weight model. Offering a 1 million token context window and a maximum output of 384K tokens, the model raises the bar in coding and agent benchmarks.
With the V4.1-Flash-1017 update to its 552-billion-parameter MoE architecture, DeepSeek lifts its Terminal-Bench 4.0 score to 42.3 and its Humanity's Last Exam result to 42.6, while offering full out-of-the-box optimization for agent frameworks such as Claude Code, DeepSeek Harness and Tencent CodeBuddy.
We compare TypeSafe AI Jev and AWS Strands Decider 2B, two key representatives of the decision models becoming the new standard in AI agents, on intelligence score, local execution, calibration and cost.
We take a detailed look at the millisecond latency advantages and zero output token cost of using TypeSafe AI's Jev decision model instead of traditional large language models in production AI workflows.
We look at how you can solve wrong function calls and broken schemas in autonomous AI agents running in production with the TypeSafe AI Jev model's zero-type-error, calibrated confidence score architecture.
We examine the technical details of the TypeSafe AI System One architecture inspired by Daniel Kahneman's cognitive theory, the RLCD training methodology and the Jev model's record 1.6% calibration accuracy.
The US government has issued a sweeping security directive banning products of China-based AI startup DeepSeek, its open-weight models and third-party services built on those models from all civilian federal agencies.
Developed by Meta, Muse Spark Open has 816 billion total and 42 billion active parameters and launches with a 1 million token context window, advanced coding skills and local execution on two M5 Ultra Mac Studios with 1 TB of unified memory.
Filing for bankruptcy just five years after its founding, Character.AI has struck a deal with Disney for a $20 million asset sale.
Google's new model Gemini 4 Argon is now available, with benchmark results that beat GPT-6 Astra and Claude Opus 5.5 and a price one-fifth of its rivals'.
Mistral AI crowned Europe's sovereign AI push with the 1-trillion-parameter Mistral Large 4. With an MoE architecture of 49 billion active parameters, a 1M context window and a plan to release open weights, a new era begins in the open-source world.
To meet European Union transparency standards, OpenAI is rolling out textGrain, a system that adds invisible watermarks to ChatGPT and Codex text.
Aleph Alpha announced Kolibri-1, an open-weight model under the Apache 2.0 license that uses only 3.46 billion active parameters per token and has a 1 million token context window.
Amazon Web Services (AWS) released Strands Decider 2B, an open-source decision model that minimizes token consumption and latency in AI agents by choosing actions directly instead of generating text.
Google unveiled Gemini 4 Argon, its new flagship model with a 1 million token output capacity and advanced cybersecurity capabilities.
Anthropic announced Claude Sonnet 5.5, the newest member of the Claude 5.5 family. 30% faster and 30% more economical, the model takes the top spot in autonomous coding with a 70.6% score on Terminal-Bench.
OpenAI introduced GPT-6.1 Sol, which offers coding and reasoning power at the level of its flagship GPT-6 Astra for one fifth of the price, along with an Ultrafast tier that generates 300 tokens per second.
EVREN, the national AI platform developed by the SSB, is now available with a high-performance Nvidia H200 GPU cluster, a pool of open-weight LLMs and an architecture built around data sovereignty.
OpenAI is getting ready to announce a $500-a-month Pro Max plan for ChatGPT, with Cerebras infrastructure, high inference speeds and autonomous tasks at the forefront.
194 articles about LLM, often together with AI, Technology, MCP. The newest one is "Anthropic Unveils Claude Fable 5.2 and Mythos 5.2: Frontier AI for Coding, Science, and Cybersecurity".