Inception ships Mercury 2.5, the largest diffusion LLM yet
Two releases in one afternoon: a startup argues that text generation doesn't have to be sequential, and a frontier lab walks out of the industry's biggest lobbying tent over chip exports.
Inception has released Mercury 2.5, which it calls the most capable diffusion language model on the market and, to its knowledge, the largest one ever trained. Diffusion LLMs are the road not taken by nearly everyone else: instead of writing one token at a time, each depending on the last, they refine a whole block of text in parallel, the way an image model resolves noise into a picture. Inception says the new model is a significant step up in quality over Mercury 2 while keeping the same low-latency, low-cost serving profile, and that it was built from the failure cases its customers hit in production rather than from benchmark chasing. Early numbers from users back the pitch: voice-agent builder OpenCall reports median model response latency near 170 milliseconds, with P99 falling from minutes to about a second, and Augment Code says moving context compaction onto Mercury cut that step from roughly 150 seconds to 27 seconds at 90 percent less cost.
The interesting part is not the raw speed — it's that Mercury 2.5 is being positioned as the cheap, fast supporting cast around a frontier model rather than a competitor to one. Coding agents already split work across models: one plans or writes code while others search, route requests, compact a long session, and summarize tool output. Those calls repeat constantly, so latency and cost compound, and nobody needs GPT-class reasoning to summarize a tool result. Inception claims quality competitive with GPT-5.6 Luna at its low setting and Gemini 3.5 Flash-Lite, with a 260K-token context window and tunable reasoning depth. Standard pricing lands at $0.20 per million input tokens and $0.75 per million output, discounted 80 percent during launch to $0.04 and $0.15. The claim to be the largest diffusion LLM ever trained is Inception's own and unverifiable from outside; what's verifiable is that a non-autoregressive architecture is now holding down real production voice and coding traffic, which is the thing skeptics said wouldn't happen.
Anthropic is ending its membership in the Information Technology Industry Council after the trade group asked Congress to strip three chip export-control measures from the annual defense bill. ITI — the lobby that speaks for much of Big Tech in Washington — sent a letter to the Senate and House Armed Services committees last week urging removal of the provisions from the National Defense Authorization Act. Anthropic supports all three, including the AI OVERWATCH Act, which would require the Commerce Department to notify Congress before approving licenses for advanced dual-use chips headed to adversarial countries, with a review period before those licenses proceed.
This is a genuinely rare split. Industry coalitions usually hold a line on trade even when members disagree at the margins, and Anthropic's exit means one of the three largest frontier labs is now lobbying against the position of the group that represents Apple, Amazon, Google, and Microsoft. It also cuts against the obvious reading of Anthropic's incentives: export restrictions can raise its own compute costs and shrink the market for its chips' suppliers, and Claude depends on enormous training and inference capacity. The company has taken the harder line on chip controls before, including supporting limits on advanced chip sales to China, on the argument that compute access is a strategic US advantage worth protecting. ITI's opposition shows that sharing a concern about regulation doesn't produce agreement on national security controls. The Senate returns from recess on September 14 with the defense bill awaiting floor action, so there's a short window to see whether the provisions survive. Whether Anthropic's departure changes anything depends on Congress, not on Anthropic — but it does give lawmakers cover to read chip exports as a national-security question rather than a commercial one.
A developer group has run Moonshot's full 2.8-trillion-parameter Kimi K3 on a MacBook Pro by streaming the model off four SSDs. The project, ARGODRIVE Deltafin, is a fork of an existing single-binary runner and makes a point of not quantizing: all 16 active experts per layer, every byte as Moonshot shipped it. That's the distinction it draws from other local K3 setups, which re-encode the expert bank down to roughly 3 bits to fit — fast, but no longer the released weights, with nobody measuring what the compression costs.
Be honest about the number, though: the full-fidelity run manages about one token per second, or roughly one per minute on an M1 Max with 64 GB, against the 16-node, 4.8-terabyte-VRAM cluster Kimi K3 was built for. That is not an interactive assistant. What it is, is a proof that an air-gapped, uncompromised copy of a frontier-class open-weight model can be executed on consumer hardware — which matters for offline audits inside closed corporate and government perimeters, where uploading a codebase to an API is a non-starter. We covered the other end of this spectrum in Microsoft's Project Zenith runs 30B models on dev PCs; Deltafin is the same instinct pushed to an absurd extreme. The open question is whether weight-streaming improves fast enough to make unmodified frontier weights practically usable locally, or whether quantization wins by default.
What to watch: whether the export-control provisions survive the NDAA floor fight, and whether any independent benchmark puts Mercury 2.5 next to the frontier models it says it can replace for the cheap calls.
Is a one-token-per-second full-fidelity local run a real capability or an expensive stunt? Tell us in the comments.
Sources: ARGODRIVE Deltafin (GitHub)
