Big milestone for Reducto: we’ve raised a $75M Series B led by a16z, bringing our total funding to $108M. Since our launch, we’ve been obsessed with pushing the limits of what’s possible when AI meets the real-world messiness of documents. We’ve now processed 1B+ pages, nearly 6x the pages since our Series A less than 6 months ago. We’re doubling down on model research, developer tooling, and the reliability that production AI demands. We'll be announcing a series of product improvements that will extend our capabilities even further, helping our customers get even closer to perfect accuracy across their most challenging docs. A big thank you to everyone who helped get us here. But we're just getting started - if you're interested in what we're doing, we’re hiring across all functions! reducto.ai/careers

deepseek posted way more alpha on their zhihu page, if you only read twitter you are reading the laowai approved stuff "Based on this framework, we have implemented a high-throughput, low-latency large-scale inference service for DeepSeek V4.1 Flash. Currently, the service throughput of a single cluster is 155 trillion tokens/day, and the overall revenue is $38 per card per hour (based on off-peak prices)." it costs roughtly $5-10 per gpu hour so depseek is claiming 80% serving margin at THEIR prices lololololol

we built classify out of the pain of seeing customers use our extract endpoint to sort documents, which is an overkill when just need to find out what each document is. classify is a smaller, purpose-built endpoint that's: > faster > costs 0.5 credits per page > has already processed nearly 1m pages. around the same time, we also rolled out batch for less time-sensitive workloads, giving parse + extract jobs a 20% credit discount. in this video, our engineering team walks through how both came directly from patterns we saw in customer workflows!

what happens in tulum stays in tulum (except the vlog)

disagree with a lot of his takes but this is a good one. a ton of the leaders in AI I’ve talked to consider gradual disempowerment/transhumanism an eventuality. it’s scary.

@OrnnExchange ?

changed the default model for my main agent thread in raubot (my personal agent) to muse glimmer! very unexpected subagents still do most of the coding and are kept as opus

muse writes so well damn

with r-1, we've replaced a long pipeline of specialized models with one model that sees the whole page. that ultimately translates into fewer handoffs where context can get lost. we invited @ani_0203_ and @chuyishang to walk through a few of the edge cases they worked on, from watermarks and handwriting to figure summaries that use the surrounding page for context. really proud of what the team built here!

I wish there was a subscription I could pay for that gave me access to try SaaS that is locked behind “Contact Sales” - really useful especially for sales demos where I could tell my agent to point out flaws within their existing product experience that could be fixed by my software

cloudflare is too good

The cambrian explosion of internal apps after releasing this should be studied. Every customer can get a unique demo page or benchmark. A dashboard for every product rollout etc. All done without compromising security. Insane work @abhiarya!!

“Why not do this for everything?” is exactly the founding question for our company We know what the verification loop is for HTML parsers, but what is it for… tax software? robotics control systems? UX researcher agents? Construction AR? niteshift.dev

Just heard from my coworkers (@jeanghislainbil and @ani_0203_ ) that r-1 is running at 500-800 decoding tps per request. On simple pages, p50 e2e latency is ~800ms. We are serving r-1 on Nvidia chips. Give it a try!

what does "good" document processing even mean? benchmarking in this space can mean a lot of things: > accuracy on complex docs > latency > throughput for scale we're not here to tell you what set of documents is a "definitive benchmark", because your own docs are the ones that matter. the more important part is actually deciding how you measure performance. @archanaahlawat and @jannalulu_ have been working a lot on building and tuning our in-house judge framework, and turns out there's a couple of reoccurring patterns we see that might help you with your eval process:

the agent/harness you build for yourself will likely always be the best one for you. was frustrated for so long that muse/instinct/etc. didn't support bitwarden, my preferred password manager. my harness now supports bitwarden passwords, TOTP, and passkeys for computer use!

I built a headless slack client for agents to triage your slack unreads for you. it streams live slack state to a durable object on cloudflare. oss! try it here: github.com/raunakdoesdev/sla…

yoooo

pi durable is so so good @badlogicgames built an agent on top of @VictorTaelin ‘s OptChat memory idea (genius!) that runs in a durable object and can hack on/redeploy itself (resuming durably without issues!) hooked it up with codemode +@RhysSullivan ‘s executor and it’s absolutely ripping don’t know why i’d bother using instinct/muse/etc. after this