Large language model (LLM) news — GPT-4, Claude, Gemini, Llama, Mistral and the latest research on training, fine-tuning, RLHF, and deployment of LLMs.
LoreDocs launched a tool that imports Obsidian vaults into a queryable SQLite database, enabling full-text and semantic search via MCP integration with Claude and other AI agents. The import preserves folder structure, Y…
Researchers Nie et al. published a paper on arXiv in 2025 introducing an 8-billion-parameter diffusion language model that generates text by un-masking tokens rather than left-to-right, rivaling autoregressive models. An…
Researchers Deng et al. published a paper on arXiv in 2025 introducing a unified multimodal pretraining model that understands, generates, and edits content, with multimodal reasoning emerging as data scales. Rudrite Res…
DeepSeek-AI published DeepSeek-V3.2, an open large language model featuring Sparse Attention that reduces computational complexity from O(L²) to O(Lk). An interactive visual explainer of the paper is available, with exhi…
DeepSeek-AI published DeepSeek-OCR: Contexts Optical Compression on arXiv in 2025, achieving ~10× context compression with 97% decode accuracy by rendering long text as images. An interactive visual explainer is now avai…
AI subscriptions priced at $20/month are heavily subsidized by venture capital and are not sustainable, with enterprise pricing already reaching $30-$100+ per user. As AI companies face IPO pressure and investor demands …
Government-imposed AI model bans are becoming more common, with countries like Italy, Australia, and South Korea restricting access to models such as ChatGPT and DeepSeek. To avoid workflow disruptions, organizations mus…
OpenRouter Fusion achieves 64.7% on key benchmarks versus Claude Fable 5's 65.3%, but at roughly half the cost per token, making it a more cost-effective option for high-volume workloads. The 0.6% performance gap is negl…
OpenRouter launched Fusion, a multi-model API that fans prompts across several AI models in parallel and synthesizes their outputs, achieving performance near Anthropic's Claude Fable 5 at roughly half the cost. The serv…
PostHog introduces 'build mode,' enabling non-engineers to ship code via its Slack app, AI, and MCP tools, with users creating over 500,000 insights and 400,000 AI conversations in 30 days. The platform allows anyone to …
LLM costs scale linearly with usage, and enterprises spending over $10,000 annually can optimize by implementing token budgets, choosing between API and local inference, and using fallback strategies. Token budgeting met…
LLM guardrails control risk by constraining model behavior through input validation, output filtering, and fact-checking. Key strategies include prompt sanitization, length limits, content filtering, response validation,…
Multi-model system design involves orchestrating multiple AI models to handle complex tasks that a single model cannot. Five key architecture patterns—single model, sequential, parallel, hierarchical, and ensemble—offer …
Haystack introduces serialization for LLM pipelines, enabling conversion to a format that can be saved, loaded, and transmitted across networks. The feature supports components like TransformersChatGenerator and ChatProm…
NVIDIA released Nemotron 3 Ultra, a 550-billion-parameter open-weights reasoning model, on June 4, 2026. It is the best US open model by Artificial Analysis's scoring but trails Chinese leader Kimi K2.6. The model is des…
The US government issued an export-control order disabling Anthropic's Fable 5 and Mythos 5 models after a jailbreak, while New York's attorney general subpoenaed OpenAI over advertising and data practices. These actions…
Researchers Chen et al. published a paper on arXiv 2025 introducing Janus-Pro, a unified multimodal model that uses separate encoders for understanding and generation within a single transformer. The model scales data an…
The Kimi Team published a new attention architecture, Kimi Linear, on arXiv in 2025. The architecture uses Kimi Delta Attention interleaved 3:1 with full attention, becoming the first linear hybrid to outperform softmax …
Researchers Zhao et al. published a paper on arXiv 2025 introducing Absolute Zero, a method where a model proposes its own tasks and a code executor grades them, enabling reasoning reinforcement learning with zero human …
OpenRouter launched Fusion API, a compound AI system that orchestrates up to eight LLMs in parallel to improve reasoning accuracy, achieving 69% on the DRACO benchmark. Microsoft released seven new MAI models built from …