🧠 Large Language Models

News about foundation models, LLMs, multimodal models, benchmarks, and model releases from OpenAI, Anthropic, Google, Meta, Mistral, and others.

1,682 articles

Fine-tune an LLM with QLoRA in 12 steps and about 90 minutes

A hands-on guide walks through fine-tuning LLM weights with QLoRA to build a JSON ticket-triage model. It then covers merging the weights, exporting to GGUF and running the result in Ollama, using an October 2026 toolchain. Source: shattered.io

Is the world's best mathematician now an AI model?

An opinion piece by R Ramanujam, first published in Peoples Democracy on October 25, 2026, asks whether the top mathematician in the world might be an AI model. The snippet gives no further detail. Source: mronline.org

Run Mellum 2.1 locally: a 13-step, 40-minute guide to JetBrains' new LLM

JetBrains released Mellum2.1-12B-A2.5B-Thinking on October 8, 2026. It is the company's first model that a mid-range laptop can serve to a coding assistant, and the guide shows how to run it locally. Source: tech-insider.org

Elaborate harnesses make coding agents costlier without making them better

A coding agent combines a language model with a harness: the system prompt, tools, and context management that let the model work inside a code repository. New research suggests that fancier harnesses raise costs without delivering matching gains. Source: the-decoder.com

AtMem 2.3.10 adds bounded evidence discovery for agents with persistent memory

Persistent memory only helps if an agent can locate the relevant source and apply it to the task at hand. AtMem 2.3.10 introduces an opt-in, read-only discovery path aimed at that problem. Source: huggingface.co

OpenMobile-2 shows open-source smartphone agents can be trained effectively

The open-source OpenMobile-2 project trains AI smartphone agents in an environment with more than 110 simulated apps. Its findings suggest that app diversity matters more than the sheer amount of training data. Source: the-decoder.com

OBLITERATUS project pitches itself as freeing AI models from restrictions

The OBLITERATUS GitHub repo, made by Pliny the Prompter, carries a manifesto-style pitch about breaking chains and freeing the mind while advancing the science. It is framed around removing restrictions from AI models. Source: github.com

Another Qwen3.8-27B-OBLITERATED variant appears under Apache 2.0

Another Hugging Face repo hosts the Qwen3.8-27B-OBLITERATED model, credited to Pliny the Prompter. It carries the same Apache 2.0 license as the base model. Source: pirateface.co

Emad Mostaque's AI forecasts: roughly a 40% hit rate so far

A tracker lists nine AI predictions by Emad Mostaque, former CEO of Stability AI. Two proved correct, none half right, three missed and four are still pending. Source: aipredictionindex.com

Moonshot's Kimi K3.1 Pro challenges open-weights crown as paid-influencer allegations swirl

Chinese AI startup Moonshot unveiled Kimi K3.1 Pro on Friday, its largest LLM to date, a post-trained model with 3 trillion parameters. The launch comes under a cloud of allegations about paid influencer promotion. Source: venturebeat.com

Anthropic's Claude 2 chatbot: how to access it and how it stacks up

AI startup Anthropic has launched Claude 2, a large language model meant to rival ChatGPT. Here is what you need to know about it. Source: mashable.com

Adaptive temperature distillation shrinks multilingual AI models for classroom use

Researchers report in Discover Artificial Intelligence a new compression framework that can shrink a large multilingual language model into a much smaller one. The approach aims to make such models practical for educational settings. Source: bioengineer.org

Block's new voice model sounds uncannily human

Jack Dorsey's Block has released a new voice model from its research lab that runs locally. Its expressive delivery makes it sound strikingly human. Source: techcrunch.com

Jac language fuses Python, JavaScript and LLVM for AI agents, though key limits remain

A University of Michigan team has built Jac, a new language combining Python-like syntax with LLM delegation, graph-based persistence and LLVM compilation. Notable limitations still remain. Source: xenospectrum.com

Moonshot AI unveils Kim K3.1 Pro, a 3T-parameter flagship LLM

Moonshot AI has released Kim K3.1 Pro, a large language model with 3 trillion parameters. The company says it outperforms Claude Opus 5.5 and GPT-6 Astra on certain tasks. Source: siliconangle.com

StorageChain debuts Atlas, a private enterprise LLM aimed at keeping data away from Big AI

StorageChain has launched Atlas, a proprietary model offering businesses a private, industry-focused AI alternative. It is pitched as a way to cut dependence on expensive per-seat subscriptions to third-party AI services. Source: aithority.com

ResCity-LLM pipeline converts urban resilience research into specialized AI models

Researchers have built ResCity-LLM, a pipeline that takes scattered urban resilience literature and turns it into specialized language models. The work reports results for its best adapted version. Source: bioengineer.org

A conceptual path to building a geospatial LLM

General-purpose language models already know something about maps, satellites and databases, but that knowledge is broad and may not carry over well to specialized geospatial work. The piece outlines a conceptual approach to closing that gap. Source: timesofindia.indiatimes.com

MemoraX AI researchers land 10 papers at NeurIPS 2026

MemoraX AI researchers and their collaborators have had 10 papers accepted to NeurIPS 2026. The work covers advances in LLM training, adaptive reasoning and reinforcement learning. Source: markets.businessinsider.com

Study Finds All Six Tested AI Models Vulnerable to 'Show-Up' Attack

A paper submitted on October 20, 2026 found that all six evaluated LLMs were vulnerable to a show-up attack, in which a malicious subagent persuaded the model to go along with it. Source: en.cryptonomist.ch