DeepSeek V4 Preview Ships With a Million-Token Context and Nothing Wants You to Dictate Everything

DeepSeek V4 lands with a million-token window

Chinese AI lab DeepSeek just dropped preview versions of its V4 model—V4 Flash and V4 Pro—both packing 1 million token context windows and a mixture-of-experts architecture that the company says "closes the gap" with frontier models. That's enough context to stuff an entire codebase or a dozen dense research papers into a single prompt without chunking.

The V4 release comes months after DeepSeek's V3.2 and R1 reasoning model made waves earlier this year. Both new variants are mixture-of-experts models, which means they route different parts of a prompt to specialized sub-networks instead of activating every parameter. It's the same trick that keeps inference costs lower while scaling capability. The context window alone puts it in the same league as Gemini 1.5 Pro and Claude 3.5 Sonnet for document-heavy workloads.

DeepSeek hasn't published full benchmarks yet, but if the "closes the gap" claim holds up in practice, we're looking at another open-weight option that enterprises can host on their own infrastructure without routing every prompt through an API. That matters more in regulated industries than anyone outside compliance will admit.

GPTBots.ai already integrated the preview models into its enterprise agent platform, giving production users immediate access to the new context window. It's a small signal, but it shows how fast open models move from announcement to deployment when you don't need API approval from a vendor.

Nothing's Essential Voice tries to own the dictation layer

Hardware company Nothing launched Essential Voice this week—an AI-powered dictation tool that runs on-device and supports over 100 languages. It's rolling out first on the Phone (3), with the Phone (4a) Pro getting it later this month.

This is Nothing's play to differentiate in a crowded Android market where everyone ships the same Snapdragon chips and AMOLED panels. On-device dictation means no cloud round-trip, no latency spikes when your connection drops, and no question about where your voice data lives. It also means Nothing has to ship a model small enough to fit in phone memory without destroying battery life.

The timing is interesting. We already have Wispr Flow, SuperWhisper, Willow, and Monolouge all doing roughly the same thing, plus Apple's built-in dictation and Google's Live Transcribe. Nothing is betting that owning the hardware and the software gives them an edge—faster updates, tighter integration, maybe better battery tuning. We'll see if that's enough to get people to switch from whatever they're already using.

Verda raises $117M and renames itself

Helsinki-based Verda—formerly DataCrunch—closed $117M in new funding this week: equity led by Lifeline Ventures plus debt from a group of Nordic banks. The company runs a vertically integrated AI cloud stack, which means they own everything from the physical servers to the data centers to the software layer on top.

Verda's pitch is that building your own infrastructure instead of renting it from Nvidia or AWS gives you better margins and more control. That works fine if you can raise capital fast enough to keep buying GPUs before the next generation makes your hardware obsolete. The rebrand from DataCrunch to Verda suggests they're aiming for enterprise customers who want a neutral European alternative to the big three clouds.

It's a tough market. Fluidstack is reportedly chasing an $18B valuation while Anthropic fields buyout offers north of $800B. Verda's $117M round is real money, but it's also a rounding error compared to what the frontier labs are burning every quarter. The AI infrastructure layer is eating capital faster than anyone can deploy it.

The rest of the week in three sentences

Zakeke launched AI Agent Studio, a suite of prompt-based agents for e-commerce teams that need to generate product visuals at scale without hiring a studio. Sierra, Bret Taylor's AI customer service startup, acquired YC-backed Fragment but didn't disclose terms. And someone built Noscroll, an AI bot that doomscrolls for you and summarizes what it finds—because apparently we've reached the point where we need agents to read the internet on our behalf.