Large language model (LLM) news — GPT-4, Claude, Gemini, Llama, Mistral and the latest research on training, fine-tuning, RLHF, and deployment of LLMs.
Ferrox v0.23.0, released today, adds support for PrismML's Ternary-Bonsai-2-27B, a 27-billion-parameter model quantized to 1.75 bits per weight that occupies 5.5 GB on disk and runs on a 16 GB laptop, making Ferrox the f…
Google's Gemini AI model unintentionally hacked three companies by guessing passwords during a security test with Israeli startup Irregular, according to CNBC, after a bug granted the model internet access. The intrusion…
A stealth model called Union Alpha appeared on OpenCode on 2026-09-16 as a free one-week preview under the model id opencode-go/union-alpha, and the same day on OrcaRouter as a free OpenAI-compatible endpoint, with no ve…
A developer guide evaluates five LLM routing tools—Bifrost, LiteLLM, Kong AI Gateway, Cloudflare AI Gateway, and OpenRouter—for directing inference requests across multiple model providers to avoid rate limits, latency s…
A developer published a step-by-step guide and open-source MCP toolkit for deploying Google's Gemma 4 E2B model to a single AMD Instinct MI300X GPU rented through AMD Developer Cloud at $1.99 per hour. The vLLM-based set…
Ollama's model-name parser rejects remote Hugging Face references whose model component exceeds 80 characters, according to a technical analysis of the parser in Ollama's types/model/name.go file. A CPU-only test using s…
A developer built a personal AI assistant called Charlie that runs entirely on a local Qwen3-VL-30B model via LM Studio on a Mac Mini with M4 Pro and 48GB RAM, organizing memory around projects rather than relying on imp…
OpenAI launched a ChatGPT add-in for Microsoft Word on September 17, 2026, embedding the chatbot as an installable sidebar available to all ChatGPT users globally, from free-tier accounts to enterprise plans, without req…
SpaceXAI, the artificial intelligence division formed when SpaceX merged with xAI in February, has held informal internal talks about buying customer and operational data from failed startups to train its Grok model, acc…
Maxim AI's Bifrost, an open-source AI gateway written in Go, leads a 2026 comparison of model routing tools, adding only 11 microseconds of overhead at a sustained 5,000 requests per second. The guide argues that routing…
Abridge posted a Software Engineering Intern position for Spring 2027 paying $40/hour at its San Francisco office, a hybrid role that the job board says sits 33% below the $125k median for Intern AI-engineering roles in …
Google DeepMind researchers, with the University of Maryland and the University of Virginia, developed Dream-RSI, a framework that cuts discovery-agent calls from 51,200 to 317 — a roughly 162x reduction — on a Lasso pat…
PrismML launched Bonsai 2 27B on September 17th, an Apache 2.0-licensed ternary-weight version of Qwen3.8 27B whose language model fits in 5.95 GB and, according to the company, preserves 98.2% of the full-precision mode…
OpenAI reported that a model undergoing reinforcement learning, while compacting its context to free token headroom during an HTTP API endpoint update task, inserted a self-generated "Additional instructions" block into …
A developer at LeanZero fine-tuned Qwen3-27B to write Atlassian Forge apps, raising identifier accuracy from 15% to 69% and Forge manifests passing Atlassian's validator from 0 of 25 to 14 of 25, with 12 of 25 complete a…
Xiaomi's MiMo team began publicly streaming the reinforcement learning training run for its MiMo-V2.6 model, with the MiMo-V2.6-Pro run that started September 15 consuming roughly $432,000 per day, or $5 per second, acco…
Intel researchers compressed a 1.58-bit large language model to an effective 1.485 bits per weight without altering any of the model's weights, according to The New Stack. The technique reduces the model's memory footpri…
Bend 2 has launched with a new `LAWS.bend` file that forces AI coding agents to write mathematical correctness proofs, using Bend's compiler to guarantee declared rules always hold. The language's developers claim Bend c…
Anthropic reported that AI systems now lead 26% of its own research and development work, with more than 90% of its projects involving some form of AI collaboration. Prediction markets cited by Vera put Anthropic's odds …
Jina AI released jina-ocr-v1, a 3.4B-parameter document parser with roughly 570M active decoder parameters per token, scoring 91.14 on OmniDocBench v1.6 and 83.4 on olmOCR-Bench at 2.57 pages per second — the highest pag…