OpenAI Consolidates Product Teams for the 'Future of Agents'

OpenAI has revealed that it is consolidating ChatGPT, the coding agent Codex, and its developer API into a single product team. The new team will be led by Codex lead Thibault Sottiaux, with the eventual goal of building a "super app" that also integrates the Atlas browser. Co-founder Greg Brockman now officially oversees overall product strategy.

This signals a clear strategic shift for OpenAI, moving from offering standalone tools toward a single platform that integrates agentic capabilities. If chat, coding, and browsing — previously separate functions — come together into one experience, user workflows could change significantly.

Mistral CEO Warns Against Scanning French Military Code With Anthropic's Mythos

Mistral CEO Arthur Mensch warned of the dangers of letting US AI models — specifically Anthropic's Mythos — scan French military codebases. He noted that modern AI can orchestrate attacks and suggest exploits, pointing to the risk of deepening security dependencies.

While acknowledging that Mistral's own models pose similar risks, Mensch emphasized the importance of Europe building its own AI infrastructure. He also denied that Mistral is up for sale and revealed plans to pursue an IPO. His remarks underscore how AI's geopolitical dimension is becoming an increasingly prominent debate in Europe.

llama.cpp Improves Quantized KV Cache Handling for Tensor Parallelism in Dual-GPU Setups

Improvements to tensor-parallel splitting in llama.cpp on dual-GPU setups are generating buzz in the LocalLLaMA community on Reddit. Until now, `--split-mode tensor` only supported a non-quantized KV cache, which had led many users to give up on tensor parallelism.

A community member has released a patched fork based on mainline. Running Qwen 3.5 27B Q4_K_M on an RTX 3060 12GB + RTX 4070 Super 12GB setup, they achieved roughly 40% faster performance — from 21.2 tok/s without tensor parallelism to 30.05 tok/s. The fork also supports the latest changes to MTP (Multi-Token Prediction).

MTP Proves Effective for Qwen 3.6-27B on AMD Strix Halo + Vulkan

Also on Reddit, benchmark results were shared for running the Qwen 3.6-27B Dense model with llama.cpp on AMD's Strix Halo APU. The comparison used the Vulkan backend while tuning MTP's `spec-draft-n-max` parameter.

Compared to 12.5 tok/s in normal execution, MTP (spec-draft-n-max 3) delivered 18.7–23.2 tok/s depending on the task, a speedup of up to about 85%. Interestingly, spec-draft-n-max 3 produced better results than 6, showing that the optimal MTP setting depends on the task.

Discussion on Using Local LLMs as Research Agents

A question posted on Reddit's LocalLLaMA about how an undergraduate student can build local LLM agents for research purposes has sparked an active discussion. The student has access to an RTX 6000 PRO and is currently running a Hermes agent based on Qwen 3.6-35B. Community members shared insights on approaches to building agents that automate paper search and research assistance.

The discussion highlights that integrating local LLMs into practical research workflows is a major area of interest within the community.

---