Inception’s Mercury 3 Brings a Diffusion Model to the Fast Tier

Ethan Mollick calls the new Claude Projects a nice interface for multi-agent work: a user-facing coordinator model and independent agents running threads that pass messages around. It “takes some fiddling but quite powerful,” he says, and has no relation to the old Projects.

@sudoingX says he is “falling for” his own agent harness. xpress agent is written in Rust, ships as one binary under 10 mb, uses 14 mb of RAM for the whole loop and finds his local model by itself. offrail 27b is the model behind it, running on his laptop. It writes its own skills as it works, and on one run it rewrote its image skill without being asked.

@MiaAI_lab shares results from @cybersec: an abliterated DeepSeek V4.1 Flash 1017 found one new vulnerability that a frontier model did not discover. @cybersec runs an operational pass against a pinned Go repo for a production app with known vulnerabilities, with Grok as the frontier baseline for all open-weight evals.

@kimmonismus flags Meta’s technical report on Muse Glimmer, a 30B open-weight model built for agentic tasks on consumer hardware. Meta says it runs at practical speeds on a Mac or PC with a single consumer GPU at 4-bit precision.

DeepSeek open-sourced DeepInfer, the inference engine behind its own production services. @MiaAI_lab lists draft tokens for faster output, custom kernels, skipping repeated work on long prompts, and the same output for the same input every time.

@sudoingX rebuilt his DeepSeek V4.1 Flash weights on the 1017 update. His v1.2 runs on 2x DGX Sparks at 141.3 tok/s on structured output, 100.7 on code and 59.2 on prose.

“Looks like today is the day,” @kimmonismus wrote on Tuesday morning, quoting a one-word post from @ChrisGPT: “Tomorrow.”

Video Rebirth open-sourced HyperFlow, a speed-up for the MiniMax H3 video model, @kimmonismus reports. The add-on keeps the same workflows and replaces the 49-step schedule with a trained 8-step one, running about 3x faster with sound generated alongside the video.

Accounting Today says AI adoption “is in the driver’s seat” while governance “is still riding in the back seat,” and looks at what the CAQ’s latest survey means for the rest of 2026. On an Accounting Today podcast, Ascend founder and CEO David Wurtzbacher suggests that getting the full capabilities of artificial intelligence may require rethinking the entire structure of the accounting firm.

The US Department of Homeland Security is ordering civilian federal agencies to remove DeepSeek and third-party products built on it, and to stop buying them, @kimmonismus reports. He reads it as a response to China’s proposed export controls on certain AI technologies, part of a broader push toward AI sovereignty.

Sam Altman described three paths to OpenAI staff, according to The Information via @kimmonismus. One is an international watchdog modeled on the atomic energy agency.

Anthropic co-founder Jack Clark went on record at the IMF about putting a ceiling on AI intelligence, The Rundown AI reports. Moderator Raghuram Rajan asked when humans hand AI the reins.

Satya Nadella says Microsoft announced new capabilities across Microsoft Sovereign Cloud in Paris, expanding choice and control for organizations running their most sensitive workloads across public and private cloud. He also met the Saint-Gobain team, which is using Microsoft Discovery to help accelerate the development of new, more efficient construction materials.