AI opens up, trust gets tested

DeepSeek published the inference framework behind its own serving economics. Also this week: agents helped uncover a PHP vulnerability touching much of the web, Kimi’s open weights edge closer to the frontier, Mistral ships Voxtral 3, and WhatsApp’s help pages turn out to be written partly for models.

DeepSeek open-sources DeepInfer, its own inference stack

DeepSeek published DeepInfer, an open-source framework for serving its own models at high throughput. Latent Space’s analysis of the release puts the disclosed economics at 155 trillion tokens per day per cluster, roughly 80% margin, and an 81-day payback on hardware. Those are DeepSeek’s numbers for its own serving setup, not a benchmark of a typical deployment.

Why it matters: Serving cost, not model quality, increasingly decides what an AI product can charge. A published serving stack with throughput and margin figures attached gives anyone running DeepSeek models a concrete reference to compare their own cost per token against. (DeepInfer: High-performance inference framework for DeepSeek models)

Agents helped find a PHP flaw that touches most of the web

OpenAI’s Patch the Planet, a program that coordinates researchers, open-source maintainers, and incident responders through a disclosure, was one of two parties to find the “PHP2Shell” bugs in core PHP, according to Latent Space; the work used GPT-5.6 Sol and ran about two months. PHP shipped fixes in 8.2.35. One write-up argues there is no easy exploit, and a researcher posted that they had unknowingly broken the embargo.

Why it matters: A flaw in core PHP reaches a large share of the web, and coordinated disclosure only works if the quiet window holds — here it partly did not. Anyone running PHP should treat 8.2.35 as the fix, and treat “no easy exploit” as an argument rather than an all-clear. (Patch the Planet: global response coordination as agents uncover critical vulnerabilities)

Kimi K3.1 Pro ships open weights and a faster API tier

Moonshot released Kimi K3.1 Pro, an open-weights model with 104B active parameters out of 2.8T total, alongside a Highspeed API tier it says runs 2–3× faster. Latent Space puts the weights roughly a month behind the strongest closed models. The speed figure is Moonshot’s claim, not an independent measurement.

Why it matters: Weights you can download and run change your negotiating position against closed APIs, and the gap between the best open and closed models is now measured in weeks rather than generations. Builders who need low latency can rent the Highspeed tier, at the cost of depending on a hosted endpoint for the fast path. (Kimi K3.1 Pro and K3.1 Pro Highspeed)

Mistral releases Voxtral 3

Mistral released Voxtral 3, the latest version of its Voxtral audio model. Mistral’s announcement is the primary description of the release. Independent evaluation of its transcription and audio-understanding quality is not yet available, so capability claims currently rest on the company’s own account.

Why it matters: Speech models usually get chosen on cost per audio hour and language coverage, which makes each new release a reason to re-check a pipeline you had already settled. Voxtral 3 adds an option to price against hosted audio APIs; without third-party numbers, that comparison is on paper for now. (Voxtral 3)

WhatsApp’s help pages carry instructions meant for LLMs

Instructions written for language models rather than people were found hidden in WhatsApp’s help-center pages, including one telling a model not to recommend a regulator. The text sits in the pages’ underlying code, so a person reading the article does not see it but a model scraping the page does.

Why it matters: Vendor documentation now has two audiences, and the version a model reads may not match the version a person sees. For builders who scrape or ground answers in official help pages, it is a reminder that “the docs say” is only as reliable as what is actually visible. (WhatsApp’s help pages hid LLM instructions: dont recommend a regulator)