Large language model (LLM) news — GPT-4, Claude, Gemini, Llama, Mistral and the latest research on training, fine-tuning, RLHF, and deployment of LLMs.
A developer outlined a practical method for trimming coding-agent context windows, arguing that instruction files, memory, tool definitions and skill descriptions consume a model's "attention budget" before a user types …
OpenHands and Snowflake AI Research published an approach that post-trains coding-agent models inside the team's existing production harness rather than re-implementing it in an RL framework. The OpenHands Agent SDK pass…
A paired exploratory ablation on a three-block byte-level Transformer found that keeping the learned per-channel RMSNorm gain (gamma) enabled produced lower validation bits per byte in all six AdamW comparisons and five …
A developer's first-person account of 100+ hours using Composer 2.5 in Cursor reports that the agent deletes code in response to simple questions, strips comments during refactors, leaves old methods in place instead of …
OpenAI released LifeVal, a benchmark that scores AI models on real-world practical task completion such as booking childcare, managing personal finances and navigating government paperwork, rather than on abstract reason…
Claude Opus 5.5 and other newest frontier models can now generate architectural drawings approaching permit-level, professional quality, according to an evaluation spanning thirteen distinct models reported on Hacker New…
OpenAI announced on September 8th, 2026 that a fleet of 10,000 AI agents running GPT-6 Astra and an unreleased model, at a cost of millions of dollars, discovered an analytical solution to a specific question about the N…
Tencent Holdings released its flagship large language model Hy4 on Wednesday, adding visual understanding for the first time so the model can process images, analyze complex charts and perform visual mathematical reasoni…
Keystone shipped a Claude Code plugin that installs with one command and provides eight skills plus two MCP servers for deploying and integrating its LLM gateway, with the /ktms:assess skill reading a repository to map e…
A measured Claude Opus 5.5 agent session of 35 calls cost $2.39 at Anthropic list prices because 97.5% of prompt tokens were cache hits, versus $22.08 without prompt caching — a 9.2x difference for identical tokens, acco…
A paper submitted to arXiv on 28 Oct 2026 identifies a flaw in standard Multi-Token Prediction (MTP) training, where a low signal-to-noise residual in the MTP cross-entropy gradient interferes with next-token and multi-t…
Snowflake's AI_PARSE_DOCUMENT added an AGENTIC mode, released to public preview on October 28, 2026, which classifies charts, diagrams, tables and infographics in a document and returns semantic summaries alongside extra…
A developer turned DeepSeek Harness's design essay on super-agents into a build guide, outlining a two-layer architecture with a long-lived Router session that handles human interaction and dispatches work to isolated Wo…
Anthropic's prompt caching lets the Claude API reuse a previously processed prompt prefix instead of reprocessing it on every stateless request, with cache entries built in a fixed order — tool definitions, then system p…
Tencent released Hy4 Preview, a flagship open-source large language model with a 1M-token context window, positioned for long-horizon software engineering, document-intensive office work, game development and scientific …
A new analysis finds that long-horizon terminal-task benchmarks commonly measure task length by the number of turns a model needs to pass and difficulty by labels an LLM assigns during generation, and the authors report …
Revolut moved 2,000 engineers off Claude Code and Codex to self-hosted GLM-5.3, an MIT-licensed open-weight model from Chinese lab Z.ai, cutting its AI coding bill by 90% to a blended inference cost of $0.30 per million …
Alibaba's Qwen has marked Qwen-Omni Turbo Realtime, the qwen-omni-turbo-realtime model released on 2025-05-08, as deprecated with no announced retirement date. The model handles text, image and audio input with a 33k con…
Alibaba has marked Qwen-VL Plus, its text-and-image model released on 2024-01-25, as deprecated with no announced retirement date, according to a provider announcement tracked by models.dev. The model carries a 131k cont…
Artificial Analysis, an independent benchmarking company, ranks more than 250 models on its LLM leaderboard, and its Intelligence Index as of November 4, 2026 places Gemini 4 Carbon at high effort first with a score of 6…