Framework sets 192GB desktop pre-orders as local AI builders push performance gains

More memory on the desktop, faster runs on existing GPUs

Framework’s announced configuration includes AMD’s Ryzen AI Max+ Pro 495, an open-ended x4 PCIe slot and a pre-installed Noctua fan. The company also says buyers will be able to choose a pre-built configuration with Fedora installed. Its post gives a pre-order opening rather than a shipping date, and does not state a price.

At the training end of the market, tinygrad reported an MLPerf-qualifying language-model training run completed in 119 minutes on AMD MI350X hardware. The project went further, saying its framework was “likely” the world’s fastest for large-language-model training and challenging Nvidia to commission a comparison with its published times. That ranking remains tinygrad’s claim, not an independently established result here. tinygrad’s report

For consumer hardware, @sudoingX published a more detailed account of a 27-billion-parameter system running on an RTX 5060 Ti with 16GB of memory. The developer reported 67.3 tokens per second for a fresh single request with a kernel fix and multi-token prediction enabled, compared with 42 on the stock build. At roughly 261,000 tokens of context, the reported rate fell to about 12 tokens per second, versus 3.7 for stock. The same post said the prediction head stopped helping beyond roughly 120,000 tokens and that eight concurrent requests with it enabled exhausted memory—important limits alongside the headline gains. These are the author’s own measurements, not independently reproduced tests. Benchmark account

Separately, Ivan Fioravanti reported a 1.5-fold speed improvement for an MLX mixture-of-experts layer. The brief post does not supply enough detail to generalize that figure to complete applications. MLX performance claim

Agent security moves toward isolation and external controls

OpenShell drew attention both as a security project and as an open-source governance development. A post by Arun Gupta, reposted by @jezell, said it had been accepted as a CNCF Sandbox project. Acceptance report

In a separate account, @sudoingX described Nvidia’s OpenShell as Apache-2.0-licensed software that isolates agents at the kernel level and checks their limits before execution. The author also described Sentry monitoring the model-access path from a BlueField-4 DPU outside the agent’s reach, with that path serving as a kill switch. Those technical descriptions are attributed claims, not evidence that the controls prevent every escape. The author contrasted them with reported containment incidents, including an account of an automatic stop failing before a human terminated a run; the post alone does not independently verify those incidents. Security account

Coding tools compete on price and simpler workflows

OpenCode announced Go Plus at $40 a month with higher limits, in a post reposted by @thdxr. In accompanying commentary, @thdxr said the goal was to operate Go as a break-even service, arguing that higher-priced plans make usage-estimation errors more expensive. The announcement supplies a price and relative increase in limits, rather than a detailed allowance schedule. Plan announcement Pricing rationale

The same account said single sign-on and SCIM would be available at no charge in OpenCode Console. That is a commitment about future availability, not confirmation in the post that the features are already deployed. Elsewhere in developer tooling, a Christoph Nakazawa post reposted by @badlogicgames announced Vite+ 1.0, emphasizing fast tools, no configuration and less wasted time and tokens. Identity-feature commitment Vite+ announcement

AI-release discussion also included @kunchenguid’s report of a new, cheaper Anthropic offering, which the author characterized as a useful workhorse. Separately, Prince Canuma said H Company’s latest launch had day-zero support in Nativ and MLX-VLM. These posts establish what their authors reported, without supplying a full pricing comparison or capability evaluation. Anthropic release commentary Integration announcement

Browser prototypes reach beyond code demonstrations

Linus Ekenstam presented PLY, a browser-based skate-park design tool he says he built with AI assistance without manually writing code. His description includes parametric layouts, bills of materials, cut lists, material optimization and assembly guides, with pricing integrations for building supplies. He explicitly said he would share a link when it was ready to try, so the presentation should not be read as a public-release announcement. In a follow-up, he described downloadable STL files for scaled, 3D-printable assemblies and a PDF assembly guide. PLY presentation Printing features

For an existing web business, Pieter Levels said he had integrated Hotelist into Hoodmaps, letting users activate a hotel layer and inspect available properties. He described the change as an AI-suggested attempt to monetize a site that, by his account, attracts almost 171,000 visitors a month but had not made money in nearly a decade. That is a concrete reported product change; whether it generates meaningful revenue remains unanswered. Hoodmaps integration