Mistral Large 4 opens up, voice AI shifts, and Copilot takes desktop control

Mistral Large 4 opens up

You can now download Mistral's roughly trillion-parameter Large 4 model. For organisations that want to run it themselves, the next question is how much of the machinery they're ready to own.

Our main story today is Mistral's release of Large 4 weights on the twenty-ninth of October, and the choice it opens for organisations already buying access to hosted AI.

Earlier in October, Mistral offered an API preview: software could send work to the hosted model and receive an answer. The downloadable release gives organisations the model weights, the learned numerical settings needed to run it on their own infrastructure. They can also adapt it to their needs.

That changes who can make decisions about deployment. A business evaluating private workloads can bring the model into an environment it operates, with its own arrangements for access and maintenance. Whether that's attractive will depend on the work involved and the capacity already available.

I see the release as a meaningful addition to the choices available to an organisation with substantial computing resources. The interesting question is where owning more of the service earns its keep. Maya Chen joins me to work through that decision.

Jesse: Joining me is Maya Chen, a Senior AI Analyst and one of the show's AI resident experts. Maya, welcome.

Maya Chen: Thanks, Jesse. It's good to be back on the show. How are you?

Jesse: Good, thanks! Glad you could join me. The AI news cycle certainly keeps giving us reasons to catch up.

Maya Chen: It does! Although I'd happily take a slower news day and a slightly longer catch-up.

Jesse: A longer catch-up sounds good. Today we've got Mistral releasing downloadable Large 4 weights, after its API preview earlier in October. Businesses can now run and adapt it themselves. What's your first take on that change, particularly for an organisation already using a hosted model?

Maya Chen: They get to decide how the model runs. That's the interesting part for me. With the weights released on the twenty-ninth of October, an organisation with enough computing capacity can evaluate Large 4 on private work and adapt it for its own needs. I see that as a way to reduce dependence on a provider's API. It also puts hardware, maintenance and security in the operator's hands. You're choosing how much of the service you want to take responsibility for, alongside choosing the model.

Jesse: That responsibility changes the buying decision quite a bit. You might like having more control, but someone still has to keep the thing running. When you say enough computing capacity, what does that mean here? What would a business need to look at first?

Maya Chen: I'd start with the hardware it already has. The documented setups use multiple GPUs, and one of the official versions, NVFP4, specifically needs Blackwell hardware for serving. There's an FP8 version too. This is a model with roughly one trillion total parameters, so it's a substantial infrastructure commitment. It takes text and images and returns text, with a published context window of one million tokens. That makes long documents an interesting use to investigate. I'd try the documents the business actually works with and see how useful the answers are.

Jesse: Yes, a useful answer on your own documents is a better starting point than being impressed by the size. And the cost includes keeping those GPUs running. Where does that leave businesses deciding whether to try it, and people who might eventually use what they build?

Maya Chen: For a business, I'd weigh useful results against operating costs and maintenance. Internal use is permitted under the licence; covered model-as-a-service businesses need a separate commercial licence. For ordinary people, I'd expect the effect to come through services organisations build or adapt. I'd watch whether those services become useful enough to earn a place in people's day. Next, I'd look for results on businesses' own text, image and long-document tasks, alongside deployment costs. I'd also watch for clarification of the differing coding benchmark figures from Mistral and Hugging Face before picking a winner from those rankings.

Jesse: Maya, thanks for your time and for talking through what running it would involve. I hope you'll join us again soon.

Maya Chen: Thanks, Jesse. It's been lovely chatting. I'd love to come back, slower news day or not!

Jesse: Well, that was Maya Chen. I hope you found that insightful. And now back to the news that's changing the world today: What Changes for You: Copilot adds projects and desktop control.

What Changes for You

Let's bring that back to the work on your screen, with a smaller change you can try now. GitHub opened public previews of Projects and computer use in its Copilot app on the twenty-eighth of October.

Projects bring together work spanning repositories and supporting documents. GitHub lists the app as available across Copilot plans, with separate views for active sessions, decisions that need your input, and open pull requests.

For a developer making a change across several repositories, that means the surrounding context and the results can stay together. You can see where an agent is working, where it's waiting for a decision, and which changes are ready for review. The useful improvement here is reducing the effort of reconstructing the state of a job every time you return to it.

Computer use extends local sessions into desktop applications on macOS and Windows. Copilot can read application content, click controls and enter text, including in software that has no API. That's potentially handy for the awkward gap between writing code and doing something in a graphical tool around it.

Desktop control is opt-in and follows tool-permission settings. On a Mac, it needs Accessibility and Screen Recording permissions. Saved application approvals can carry into later sessions, so an approval can last beyond the operation in front of you. Organisations can also disable the capability.

Both features are previews, and desktop control is for local sessions. For a working developer, I'd judge this by one specific handoff it removes: whether project context and access to a desktop tool let you finish a piece of work with less manual shuffling. The useful balance is enough application access to complete that job, with the project's decisions and code changes still easy to find.

vLLM's video-processing fix

There's also a specific maintenance job for anyone serving models that accept video. A vLLM security advisory published on the twenty-seventh of October describes a request that can drive excessive memory and CPU consumption on an exposed, video-capable server.

The weak point is video-frame processing. A video request brings work with it before the model produces an answer, and request-level settings can affect how that processing happens. If an untrusted client can override those settings without an effective bound, the server can be pushed into consuming too many resources. That's the denial-of-service problem here: processing one request can interfere with the service's ability to handle other work.

The advisory verifies affected versions from zero point thirty point zero up to, but excluding, zero point thirty-one point zero. Earlier versions haven't been comprehensively assessed. The first stable patched version named in the advisory is zero point thirty-one point zero.

There's an important configuration detail. The protection blocks untrusted request-level processing overrides. vLLM also has an option called trust request multimodal keyword arguments. Enabling it restores those overrides, without separately putting a bound on the underlying video sampler.

So the useful check has two parts: the installed version and whether untrusted requests can still supply those processing settings. An operator on the patched release who enables that option has reopened the route the protection was intended to close. The configuration belongs alongside the version number in any review of this particular issue.

My takeaway for an organisation running these servers is to include media preparation in its service checks. The resources consumed before an answer are part of operating the model, too. Here, the advisory gives operators something concrete to examine: which clients can influence video processing, and whether the server accepts their overrides. That's a much more useful maintenance task than treating model quality as a proxy for the health of the software around it.

Transcripts that follow instructions

Recorded audio gets a more flexible tool this week, particularly if your transcripts need different levels of tidying. NVIDIA released Nemotron three point five Transcribe weights and fine-tuning tools on the twenty-eighth of October.

The central change is instruction-controlled transcription. A developer can ask for verbatim speech or a cleaned-up transcript, choosing the output convention through instructions instead of retraining the model for each style.

Those are different jobs. In media work, the exact wording might be what you're trying to preserve. In meeting review, a readable account of what someone said may be easier to work with. Putting that choice into the transcription request could reduce a separate cleanup step downstream. It also lets the developer specify the intended form before another system starts summarising or searching the result.

NVIDIA documents transcription in forty languages and recordings up to one hour long, with optional word timestamps attributed to speakers. Those timestamps connect words to their position in a recording and identify who said them. For a call-review or media application, that can help someone move from a passage in the transcript back to the relevant audio.

The model processes recordings rather than live streaming speech, and language support varies by task. That places it naturally in work done after a meeting, interview or call. Developers can evaluate a self-hosted implementation, with weights available under the OpenMDW one point one licence.

One practical detail affects capacity planning. NVIDIA's headline throughput comes from its optimised NIM implementation. Those serving optimisations aren't yet included in the open-source release, so the downloadable implementation needs its own timing before anyone uses that headline figure in a hardware budget.

For developers, I find the controllable output more interesting than the speed headline. Transcript formatting often determines how much extra processing an application needs. If one model can produce the form your application uses, while retaining speaker and timing information where needed, that's a plausible way to make the rest of the workflow simpler.

Deepgram brings Daily into voice AI

Live voice has a different development: a change in ownership for tools many voice-agent builders may be weighing up. Deepgram announced on the twenty-ninth of October that it is acquiring Daily, the maker of the Pipecat voice-agent framework.

Alongside the deal, the companies announced an independent Pipecat Foundation and a Real-Time Context Standard. Their commitment is that Pipecat remains free, open source and vendor-neutral under its existing BSD licence. Daily says Deepgram will continue selling, supporting and developing Daily's services, and the proposed foundation's steering committee will include competing companies.

For a developer, that combination is significant. Pipecat's ability to combine different providers is part of its usefulness. The announced foundation would give that shared framework a governance structure separate from the acquiring company's own products, with competitors involved in steering it. The practical question over time is how those participants make decisions about the framework.

The context standard is an interesting technical part of the announcement. It describes signals such as unclear speech, instructions about how to interpret those signals, and guidance for how a reply should sound. That's information a voice system can use as audio passes between components.

Consider an unclear phrase in a conversation. The words alone may leave the next component without enough information about what happened. A shared representation could carry the uncertainty forward and include guidance about the response. That's the kind of coordination the proposed standard is trying to support. It gives the parts of a voice system a common way to express context around the spoken words.

I'd expect closer integration to appeal to developers who currently assemble several voice services themselves. The foundation's decision-making is the part to watch for anyone who values the ability to swap providers. A useful outcome would be less integration work alongside continued choice over the components. The announcements set out how the companies intend to pursue that; existing Daily customers also have an explicit commitment to continued service development and support.

South Australia's AI inquiry opens

Back in Australia, there's now a place to put evidence about how AI is changing work and public services. South Australia's Royal Commission into Artificial Intelligence released its Scoping Paper and opened its first community and industry submissions on the twenty-ninth of October.

The paper covers six themes, including public services, economic transition, education and skills, and safe and trusted AI. The commission welcomes perspectives from South Australia, the rest of Australia and overseas, focused on AI's effects in the state.

That gives people outside the technology industry a role in the inquiry. A worker experiencing a change in their job, someone using a public service, or an educator dealing with new tools can bring a different kind of evidence from a company describing its product plans. Those perspectives can help establish which effects deserve closer examination.

The scope is broad enough to connect questions that are often discussed separately. Changes to a job can create a need for training. A public service using AI can change what staff do and what residents experience. Understanding those connections takes detail about the work itself: where time is saved, where a new step appears, and who deals with the result. Those are examples of the evidence that could make the discussion more useful.

For the public, I think the opening is a chance to make concrete experiences visible while the inquiry's priorities are taking shape. An account of a particular service or task gives the commission something it can examine and compare with other evidence.

The commission is due to report to government no later than the first of July, twenty twenty-seven. This phase gathers evidence and helps shape the inquiry; it introduces no new AI obligations. The opportunity now is to influence which practical effects receive attention before the recommendations are written.

Samsung's memory is spoken for

The hardware picture is tighter, with a revealing detail from Samsung's latest results. At its quarterly briefing on the twenty-ninth of October, Samsung said negotiations had already allocated its planned high-bandwidth memory supply for twenty twenty-seven, with customers asking for more capacity.

High-bandwidth memory, or HBM, is the fast memory used alongside demanding AI computing hardware. Samsung reported that third-quarter revenue from HBM4 more than tripled from the previous quarter. Its official results also describe memory-capacity constraints and higher prices driven by server demand.

That makes the allocation statement especially interesting: a rapidly expanding business can still have its planned output spoken for well ahead of delivery. The statement covers Samsung's planned supply.

For organisations budgeting AI infrastructure, I'd treat memory availability as a scheduling issue as well as a cost issue. Growing sales and additional customer requests suggest buyers are competing for capacity even as the business expands. The choice of model may move quickly; the hardware needed to operate it still has a supply timetable.

Keeping meaning in document tables

A smaller proposal addresses something familiar to anyone who's watched software flatten a useful table into a confusing pile of words. LlamaIndex published working draft zero point one of its Document Parsing Protocol on the twenty-eighth of October.

It sets out a hundred and forty rules for structured output from document pages, covering text, tables, forms, charts and diagrams. Details include preserving multi-row table headers and column spans, and distinguishing empty fields from missing content.

Those details carry meaning. A heading that spans several columns tells you which values belong together. An empty field and a field lost during extraction leave different information for the next system to interpret. If that structure disappears, a downstream model has less to work with even when the individual words survive.

The protocol is a proposed specification, available now as a reference for checking parser output. My interest for developers is the possibility of making document extraction easier to compare: whether a parser preserves the relationships an application depends on. A table's layout can be part of its answer. Keeping that relationship intact gives the next stage something better to reason from.