Large language model (LLM) news — GPT-4, Claude, Gemini, Llama, Mistral and the latest research on training, fine-tuning, RLHF, and deployment of LLMs.
Ricursive Intelligence released Ripple-65B Preview on November 4th, a 65-billion-parameter dense open-weight coding model that the Palo Alto company says scored 80.5 on TerminalBench 2.1 and cost $32 million to train. Co…
Multiple independent groups have replicated the claim that frontier LLMs can read and write neural network weights directly without gradient training, including Hugging Face, which reproduced the MIT paper with its open-…
A benchmark called AbMaze, built by Dominik Scherm and published November 4, 2026, found that no frontier AI model outperforms a scripted Bayesian optimizer at navigating a fully measured antibody fitness landscape of 65…
Nodus published a worked analysis of the compute required to run reinforcement-learning training for LLM agents, showing that GPU cost is driven by idle time while tool calls execute in CPU sandboxes. In its example, 128…
DeepSeek open-sourced DeepInfer, the production inference engine behind its API that processes 155 trillion tokens per day, under the Apache 2.0 license. DeepInfer runs only DeepSeek-V4.1-Flash and requires NVIDIA SM100 …
Anthropic is extending text watermarking to four additional Claude models ahead of the EU AI Act's December deadline, but on older models including Claude Opus 4.6 and Sonnet 4.6, API users can disable the watermark by c…
A developer building a LangChain-based survey-report agent found that Anthropic's Claude Haiku 5.5 generated 17 bad citations in total by truncating 36-character respondent UUIDs to their first 8 characters, a failure th…
Ricursive Intelligence released Ripple-65B Preview, a 65B-parameter open-weight model it says reaches U.S. open-weight state of the art on agentic coding with an 80.52 score on TerminalBench 2.1, ahead of Reflection Beam…
A developer built Quasnooze, an open-source local AI assistant that audits each model reply against the turn's actual tool trace and memory facts rather than trusting the prose. The guard module, src/guard.js, is a pure …
A developer released Sentinel AI Gateway, an open-source proxy that sits between internal applications and external LLM providers such as OpenAI, Anthropic and AWS Bedrock to scrub personally identifiable information and…
An Nvidia paper titled "You don't need long context," authored by Peter Belcak, Pavlo Molchanov and nine co-authors, found that AI agents can match full-history accuracy by compacting their conversation history into shor…
FedRAMP granted its first "Class C" certification under its 20x revamp to Perplexity, making the company's Enterprise platform the first cloud service offering fully compliant with the new moderate-impact rules, FedRAMP …
A reproducible benchmark on identical Google Cloud c3-standard-4 hardware found that the Go-based KTMS LLM gateway added less latency, used less memory and CPU, and started faster than LiteLLM Proxy v1.106.0 running on P…
Blatent found that across six established evaluation suites, models that extracted correct answers from source text still lost credit when their wording differed from the reference, including a case where a model answeri…
Unsloth co-founder Daniel Han released Dynamic v3.1 GGUFs for Qwen3.8-Flash-Next on November 4th, claiming the quantized files score 10-15% higher than GSQ-RCO on DeepSWE, Terminal Bench, Aider Polyglot and other tests i…
Dharma AI released the Dharma Field Extractor, a 2-billion-parameter model that takes a document image plus either an open request or an exact schema and returns structured JSON with typed, nested fields, which the compa…
Google DeepMind's Gemini 4 Carbon (High) debuted at #2 on the arena.ai-run Agent Arena leaderboard on October 28, 2026, after over 21,900 sessions with a 16.2% net improvement, at a median cost of $1.80 per task — 72% le…
A large language model identified the purpose of four small Atari-playing neural networks from their raw weights alone and altered roughly 0.016% of Qwen3.5-0.8B's weights to raise its MATH-500 accuracy from 7.9% to 48.1…
A developer built LLM-Router, an open-source Rust binary that routes LLM requests across three home-network machines — a home server plus two gaming desktops — based on LLM load and server availability, with a preference…
Google launched voice remixing for Gemini 3.8 Flash TTS, available through Google AI Studio and the Gemini API, adding prompt-based editing of voice attributes including timbre, pitch, pace and accent. Gemini 3.8, Google…