Building @river_ai_inc Prev: @xai, @snap
Palo Alto, CA
Joined July 2015
- Tweets18
- Following19
- Followers123
- Likes14
We just shipped something pretty unique among post-training API providers: end-to-end custom losses.
The objective you define locally gets traced into a graph, compiled and executed efficiently on our GPUs.
Read more about the engineering behind it in our latest blog post.
Custom losses are now live on the River API.
Write your objective in PyTorch. Our Symbolic Engine compiles and runs it end-to-end across GPUs, handling efficient execution and memory usage.
Learn more and see it in action with decision models and score centering: river.ai/custom-losses
The loss is automatically chunked based on estimated memory usage and tensor lifetimes, and executed in tiles, with tile summaries checkpointed and combined in finalization.
Daniil Ostashev retweeted
We've partnered with @NVIDIAAI to make Nemotron 3.5 Super VL available on the River API from day zero.
Even before training, Super VL outperforms Kimi K2.6 on our text-to-SQL evaluation, despite being a fraction of its size. After training with our text-to-SQL recipe on River, it outperformed every proprietary model we evaluated.
These results show what's possible with River: frontier-level performance on a specialized task, using a model small enough to deploy on your own hardware.
Decided to jump on the square packing train.
So far broke three records: n=155, n=156 and n=211.
My more interesting ideas haven't worked yet. The 3 records were produced with rigorous search over corner structures, combined with soft-to-rigid like refinement.
n=67 stays unbeaten :/
ostashev.ai/square-packing
A few more interesting square packing results came from trying to generalize the problem to rectangles: let r(n, k) be the smallest h such that n unit squares fit in a h x (h + k) rectangle. With integer k these become composable into five new record blockwise square packings.
Additionally, r(163,3) produces a new record for s(129) if you cut off a few layers on the side.
We’ve just launched inference for DeepSeek V4.1 Flash and GLM 5.3 Flash for 20% cheaper than official providers.
Made possible with optimizations across the board from a custom serving stack written in Rust to a bunch of kernel magic at the lower level.
Give it a a try with River API!
We beat GPT-6 Astra and Opus 5.5 with <1% cost on text-to-SQL using fine-tuned open models.
While this result might seem crazy, this is actually the case for many (possibly all!) domain specific tasks. If you have the data – customized open models can and will perform better than closed frontier. Give it a try with River!
Introducing River Recipes: our open-source initiative for turning promising research into vetted, reproducible training recipes.
Our first recipe tackles text-to-SQL, where our trained open models beat GPT-6 Astra Pro and Claude Opus 5.5 on Arcwise-Plat for less than 1% of the cost.
Read the full breakdown and explore the recipe ↓
river.ai/achieving-frontier-…
Daniil Ostashev retweeted
River AI is growing and we've published new roles in research, engineering, and operations. Check out our careers site or reach out to our team today. If you want to build the future of personal AI, join us!
river.ai/careers
Daniil Ostashev retweeted
Today, we're sharing that River AI has raised $1.1 billion, led by @generalcatalyst and @amppublic with strategic investment from @nvidia and @AMD. Additional investors include @ycombinator and @Temasek.
We imagine a future where your AI works entirely for you and deeply aligns with your values. Compared to the corporate chatbots of today, it will feel completely different: responsible, curious, and truly yours. Most importantly, you'll own everything: the hardware it runs on, the data it learns from, and the intelligence itself.
Our training API is already live and powering incredible work today. This new funding accelerates our push to bring the full personal AI stack to life.
Read River AI CEO Igor Babuschkin’s (@ibab) conversation with the New York Times to learn more: nytimes.com/2026/08/11/techn…
Daniil Ostashev retweeted
🚀🚀🚀Omni-ID: Holistic Identity Representation Designed for Generative Tasks
snap-research.github.io/Omni…
Excited to share Omni-ID, a novel facial representation tailored for generative tasks! It captures diverse expressions & poses, enabling high-fidelity personalized generation.
Happy to have been a part of this amazing work!
My #BobaWithMoA 🧋
💫 Happy to share MoA: a *Mixture-of-Attention* architecture for personalization of generative models!
Imagine YOU🫵🏻 drinking boba🧋in a night market, or scuba diving🤿, and more!
📕: arxiv.org/abs/2404.11565
🔗: snap-research.github.io/mixt…
🎥: youtu.be/Ckv5ttqGJNA