Baseten

We’re thrilled to announce that we have raised $300M at a $5B valuation. The round is led by IVP and CapitalG, both doubling down on their investment in Baseten, and joined by 01A, Altimeter, Battery Ventures, BOND, BoxGroup, Blackbird Ventures, Conviction, Greylock, and NVIDIA. Read more here: baseten.co/blog/announcin…

Nearly half of all U.S. physicians rely on OpenEvidence every single day - at the bedside, in the operating room, and at the point of care. It surfaces research in real time, exactly when doctors need it most. Baseten powers the inference that makes this possible. Thank you, Team @EvidenceOpen, for what you do and for the trust.

OpenEvidence has become the default medical knowledge platform for over 40% of U.S. physicians; it's relied on daily for the highest-stakes decisions in medicine. Baseten is honored to power the inference behind it.

@fareehasala you can just do things 🍦

they inference engineered my ice cream???

Over 1 million clinical questions hit OpenEvidence every day. More than half the practicing physicians in the US rely on us at the point of care, mid-decision, with a patient in front of them. Downtime in that moment has real consequences. We partner with @baseten for our inference infrastructure to make sure answers are always there when physicians need them. They stopped by our office to talk about what that looks like under the hood.

@_philschmid 🤩

@baseten LGTM!

Gemma 4 is live on Baseten and available to all customers on day 0 via the Baseten model library. All models in the Gemma 4 family are multimodal, supporting text and image inputs with text output. Key capabilities include: -> Advanced reasoning and thinking -> Coding and function calling -> OCR for document understanding -> Long context windows up to 256K tokens But the most impressive is how Gemma 4 is pushing the boundaries of model architecture with innovations including alternative attention mechanisms, Proportional RoPE, Per-Layer Embeddings (PLE), KV-Cache Sharing, native aspect ratio handling for vision, and a smaller frame window for audio. All are designed to improve efficiency, accuracy, and scalability. Try it today: baseten.co/library/publis…

Our last Women in AI event might just convince me to move to NYC... I was the solo-female-engineer during all of my engineering career (similar story in academia). Being surrounded by so many excellent engineers and technical leaders—100% women—was a treat. 🩷 If you're in NYC, we're doing it again at Le Labo soon. DM me if you want to join!

@nvidia Thanks for having us!

Delivered performance, not peak chip specifications, drives AI factory productivity. Rigorous benchmarks are the only way to see past the noise. In MLPerf Inference v6.0, NVIDIA extreme co-design delivered the highest token output across the broadest range of models and scenarios. Maximizing token output drives down token cost and maximizes AI factory productivity. Read the blog post to dive into the details: nvda.ws/41aqALX @Baseten, @CoreWeave, @mlcommons

@antonpme Hi Anton - could you DM the email associated with your account so we can locate the support records and better investigate?

@baseten I already emailed you about this, but my main question is: what was I charged for? I haven't had any activity since I created the account. I’ve never used anything, and when I checked the dashboard and all the stats, there are zero projects.

🚨 @baseten is charging users for accounts with zero usage. No deployments, no requests, no nothing. The monthly bill anyway for something I didn't even subscribe to. Asked support to stop last month. Got a refund. This month: charged again. WTF? Wild business model: bill people for services they never used, hope they don't notice. 🤬

Long horizon agentic workflows will require new forms of memory that aren't just markdown memory files; this work is a meaningful step in that direction.

Pretraining is data-inefficient. This is entirely a consequence of the fact that we throw away the KV cache after every forward-backward step! If we can integrate efficient KV cache compaction into pretraining, we will unlock human level data efficiency. Neural KV cache compaction makes this possible.

What if LLMs could remember as humans do? LLM memory is either perfect and lossless or ultra-compressed. What does a slightly compressed working memory to extend its context window look like? Our researchers built a 7M-parameter perceiver that compresses KV caches 8x while retaining 90%+ factual retention. Unlike existing compaction methods, we trained a model to do this in a single forward pass. We see this as the first step toward models that actually learn from experience.

@rapprach can't wait for the Baseten-branded clothing line

Met our mascot at KubeCon 🩷💚

@tsriram @philipkiely 👀

Update: @philipkiely from @Baseten is joining us for a lightning talk at Agents and Money. The lineup keeps on getting better! If you're building agents that need to be cost-efficient at scale, this is the room to be in, with the best in the game. April 15 | Fort Mason | 5 PM RSVP → luma.com/5nqllmul

We've had a great month of March! A brief recap: -> NVIDIA GTC, featuring books, ice cream, and swag -> KubeCon EMEA, including a 2000+ person House of Kube event -> AI engineering leaders dinners at Wolfsbane (SF) and Manhatta (NYC) -> Baseten-branded ice cream social in SF -> AI/ML trivia night in the West Village -> NYC office warming party Want to attend our next event? Sign up here baseten.co/resources/type…

@BatteryVentures @Gong_io @ClickHouseDB @databricks proud to be featured alongside some of our amazing customers 🤩

Congratulations to the 4 Battery portfolio companies recognized on the 2026 #ET30 list! Late: @baseten @Gong_io Giga: @ClickHouseDB @databricks The AI supercycle is giving rise to an entirely new stack, and we're proud to back the founders building across it💥

Feel like this thesis gets more and more evidence behind it every day. Cursor, Chroma, Pinterest, Cognition, Decagon, Hippocratic, Intercom (and many many more behind the scenes) all realising that the way to own the compounding flywheel of value is specialising an open-source model to be incredibly good at the tasks you care about, with real data that continuously makes it better (which is the advantage you have over the big labs who will keep trying to eat your lunch with general intelligence improvements). my bet is that this will be the norm rather than the exception on a much shorter timeline than you think. Most companies worth their salt don't want to be beholden to or held hostage by a big lab. If you own your model, there's a huge surface area of things you can do on both the training and inference side to deliver the best possible experience to your end users