Large language model (LLM) news — GPT-4, Claude, Gemini, Llama, Mistral and the latest research on training, fine-tuning, RLHF, and deployment of LLMs.
The Modelflare team published a production guide for load testing LLM APIs that separates evidence from assumption, defining a six-checkpoint review record covering workload, traffic, routing, measurement, saturation, an…
A developer built a retrieval-augmented generation service with hybrid dense+BM25 retrieval, a 0.15 relevance threshold meant to trigger refusal, and an extractive mock generator that quotes cited chunks, then evaluated …
Google will double Gemini 4 Argon API prices on December 1, raising input tokens from $2 to $4 per million and output, including thinking tokens, from $10 to $20 per million, per the model's official documentation. Argon…
Mistral released the open weights of Mistral Large 4, a roughly 1-trillion-parameter natively multimodal Mixture-of-Experts model with a 1M-token context, on Hugging Face on October 29, 2026. The release ships under the …
An engineer published an architectural blueprint for an automated pre-production evaluation harness that uses Amazon Bedrock, telemetry tracing, and LLM-as-a-Judge quality gates to catch AI contact center agents that vio…
Mercor routed 18 million requests and 63 billion tokens through Vercel AI Gateway in its first 10 days on the platform without hitting rate limits, cutting time-sensitive batch jobs from more than 12 hours to 2 to 3 hour…
A DGX Spark user running the granite-4.2-30b-fp8 model reported getting only 4-6 output tokens per second with llama.cpp and vLLM, and reached 40 tokens per second after a forum respondent identified as lambda advised sw…
Picos is a new agent harness that gives large language models a custom 32-bit CPU called smol32 with only 32 instructions (0x00–0x1f), 16 registers (r0–r15), a single flat RAM array, raw disk storage, and a 256 × 192 RGB…
A Stanford University study published November 2, 2026 found that 44% of 50 widely cited LLM findings no longer hold when retested on 20 models released from July 2024 to September 2026 across eight developers, with anot…
A developer published an architectural guide for building production-grade Model Context Protocol (MCP) servers in pure Python using asyncio, json, and Pydantic V2. The guide covers JSON-RPC 2.0 wire format, Pydantic dat…
Researchers introduced MAServeBench, a benchmark for evaluating multi-agent LLM serving systems under hierarchical collaboration workloads, addressing how interdependent model and tool operations become ready at differen…
A research analysis argues that multimodal large language models' multi-step reasoning failures, such as losing track of which objects have already been counted, stem from a "reference gap" rather than perception errors,…
Simon Willison released llm 0.37 on 2 November, adding the gpt-6.1-sol and gpt-6.1-terra models, an `llm openai endpoint add` command for arbitrary OpenAI-compatible endpoints, and a switch of the default model to gpt-6-…
OpenAI said on November 2 that Cathedral Systems researchers will join its next internal security sprint, using OpenAI models alongside OpenAI teams to test the company's defenses, according to OpenAI's security sprint p…
Researchers introduced DeFT-RAG, a retrieval-augmented generation method that decomposes multi-hop questions into self-contained sub-questions whose answers can be reused, targeting the failure of standard RAG when facts…
A method called DecisionRank applies on-policy distillation to transfer a generative LLM's ranking knowledge into a decision-model reranker, addressing the difficulty of teaching decision models that predict choices or s…
A field engineer reports that replacing six separate notations (Makefiles, Gherkin, EBNF, init.d scripts, and others) with a single Smalltalk-80 keyword-message grammar produced a 100% pass rate (120/120) in a 360-trial …
An MIT CSAIL team reports early experiments in which GPT-6 Astra interpreted and modified neural-network weights without writing a training loop, with the authors saying one Qwen test showed a small weight change sharply…
DeepSeek's Harness team announced Creator Mode, a feature that lets users describe a plugin in plain English in chat and receive a ready-to-use extension for the DeepSeek Harness agent environment, where tools, skills an…
Alex Krentsel reproduced MIT CSAIL's Breakout weight-space programming result and found that only the latest model versions can write all 20,486 weights of the small neural network in a single reply, a capability he desc…