A large-scale open-weights release, agent security, and the philosophy behind model design — today brought movement on both the implementation and the ideas front. And once again, there is a significant update on Meta's Muse.

Xiaomi's MiMo-V2.6-Flash-RL — A 309B-Class MoE in Open Weights, Built for Self-Improvement via RL Scaling

Xiaomi's MiMo team has released MiMo-V2.6-Flash-RL, an efficiency-focused checkpoint of the MiMo-V2.6 series, on Hugging Face, along with a technical report. Built on a sparse MoE architecture, it has 309B total parameters with 15B active and a context length of 1 million tokens. It is an omni-modal model that handles text, images, video, and audio in a single model, and it also packs multi-token prediction (MTP) layers for faster inference.

The centerpiece is a self-improvement design built on "reinforcement learning scaling." It introduces numerous proprietary mechanisms, including "You Only RL Once," which trains coding, general agents, vision, and cybersecurity together in a single mixed-RL run, and Groupwise Agentic Grading (GRS/GAR), which refines the reward signal itself by comparing rollouts within the same group. On benchmarks, it posted scores that close in on frontier models, such as 67.9 on DeepSWE v1.1 and 95.1 on CyberGym.

Alongside it, the team also released MiMo-V2.6-Distill-Qwen-9B, a distillation of MiMo-V2.6 into Qwen 9B. On Reddit, reactions spread with "That's serious," and hands-on reports of the Pro version followed, such as "it needs corrections and steering, but it produces output on par with GPT-5.6 Sol." The open-weights camp keeps getting steadily deeper.

A Critical Zero-Day in Meta's Muse — A Structural Problem That "Rolls Back" macOS Defenses (Updated from Last Time)

Ars Technica reported that a zero-day vulnerability has been discovered in Meta's AI assistant Muse. Last time, the story was "an early pace surpassing ChatGPT" and "Amazon blocking its use"; this time, it is a deeply serious security topic.

Muse runs as a macOS app, and users grant it access to WhatsApp, email, calendars, and social media, as well as permissions for OS-protected resources such as file writes, the microphone, the camera, and location data. Granting those permissions to Muse effectively disables, all at once, the defenses Apple has built over the years to protect these resources from apps and terminal commands — that is the crux of the problem. Exploiting the zero-day reportedly allows apps or terminal commands running locally to completely hijack the agent, with a simple ClickFix attack said to be just one example. The gap between Zuckerberg's pitch that Muse is "built from the ground up with privacy and security" and this reality looks likely to be called into question. As of the reporting, the status of a fix has not been disclosed.

"RLCD" in the Words of Jev's Developer — A "Third North Star" to Replace Both RLHF and RLVR

TypeSafe AI's Jev has been dominating the AI conversation since its release last week. Its lead developer, Diogo Almeida (co-author of the InstructGPT paper), spoke at length about the design philosophy on the Latent Space podcast. The launch video has reportedly drawn around 40 million views, and the Discord community has grown past 100,000 members.

The core claim is "Reinforcement Learning for Calibrated Decisions" (RLCD). The argument goes: RLHF, trained on human evaluations, breeds hallucination and sycophancy; RLVR, trained on verifiable rewards, worsens "jagged intelligence" — so we should instead optimize for answers that carry "epistemically honest probabilities" on System One tasks. The observation that the mode collapse RLHF induces in probability distributions poisons calibration is a point that has rarely surfaced before.

He also offered that "the bitterest lesson is that the most important part of ML is not ML — the right task and the right data beat compute," positioning TypeSafe as "a data lab, not a model lab." He is skeptical of public benchmarks and advocates evaluating on real-world workflows.

phantom-kv — "Injecting" a KV Cache Without Touching Weights to Toggle LLM Refusal Behavior

On Reddit's LocalLLaMA, "phantom-kv," which controls an LLM's refusal behavior while leaving weights untouched, is making waves. The mechanism feeds a roughly 18MB bank of trained KV tensors into the model's KV cache as if it were "conversation history that had existed from the start." Unlike abliteration, which rewrites weights, unloading it returns the model to a byte-for-byte identical copy of the original.

Particularly interesting is the operational design: it ships with a "blue pill" mode that permits unlocking only for legitimate security work such as malware analysis, and it is designed as a "capability mode" that can be toggled on and off per request. The developer's own audit results are published alongside, openly stating limitations such as the effect thinning out after roughly 2,000–4,000 tokens in long sessions (workable via re-injection).

An Open-Source Autonomous Agent on a Long-Run Attempt at an Unsolved Math Problem — 52 Hours and 100 Million Tokens, Live

Also from Reddit. An experiment has been made public in which an open-source autonomous agent built on Qwen 3.8 27B takes on improving "C(25,15,5)", a covering design that remains an open problem in combinatorics. The best known solution stands at 42 groups, and the target is an improvement to 41. The agent has run continuously for more than 52 hours, processing over 100 million tokens, with its best result reaching 44 groups.

The agent's internal reasoning, memory, and attempt history can be browsed in real time on the site, and the correctness of solutions can be machine-verified by an independent checker. The experiment is reproducible on consumer GPUs (an RTX 5090, or an RTX 3090 depending on conditions), and it is drawing attention as a realistic step toward the question of "can open-source agents achieve research breakthroughs in mathematics?"

Policy Signals Around AI Regulation — Finland's Presidential Office Calls for Control of Frontier AI as the US Administration Rejects a "Liability Shield"

The Office of the President of Finland has published a statement titled "A Call for Control of Frontier AI Models," urging stronger controls on advanced models. In the United States, meanwhile, Treasury Secretary Bessent said in a CNBC interview that the Trump administration has no intention of giving AI industry leaders a "liability shield." The message that an administration which appears to put innovation first will not go as far as granting immunity is a signal the industry would do well to keep watching.

From Japan: Kewpie Cuts Factory Simulation Man-Hours by 80%, While "Over 70% of Mainframe Exits Fail"

Two practical stories from Japanese companies. Kewpie reports that it cut the man-hours for developing factory simulation models by 80% using generative AI. It has published an account of how it overcame the "three walls" of building such models — work that requires knowledge of dedicated software and specialized languages and tends to depend on a handful of experts.

Meanwhile, in a ranking of system-renewal articles for the first half of fiscal 2026, one forecast drew attention: "over 70% of mainframe modernization projects fail, with overconfidence in generative AI in the background." It pours cold water on the expectation that "legacy modernization gets easier with generative AI," and it is being read alongside the debate over what roles remain for people in the AI era.

Wrap-Up

MiMo-V2.6 pursuing self-improvement, a Muse zero-day born of concentrated permissions, and Jev's RLCD rethinking the north star of training. At every layer — models, security, and philosophy — the "given" is being shaken. The open-source side, for its part, keeps stacking up grounded if unflashy efforts such as the long-run attempt at an unsolved math problem, and this current looks set to continue.