A 3D cube running at 60 FPS on a Esp32 s3 microcontroller—built with TypeScript, JSX, and CSS. Very impressive.

Jev is seriously impressive. SERV makes it better. We put Jev through the same benchmark we use to evaluate leading AI models, testing it both standalone and armed with SERV Reasoning. Paired with SERV, Jev moves into the top tier of our benchmark, beating Claude Fable 5 at 30x lower cost, with cost similar advantages over GPT-5.5 (~15–20x), Gemini 3.5 Flash (~15x), and Grok 4.3 (~6–7x). The result is clear: Jev + SERV materially outperformed Jev on its own. Jev is extremely fast, well designed, and great at decision-making. Instead of relying on open-ended responses, it produces structured decisions with probability scores attached. SERV pushes that performance further, adding a stronger reasoning layer to improve the quality of those decisions for high volume agentic work. The SERV Reasoning API is live, so any developer can test and deploy SERV Reasoning today in just a couple minutes. Get access at: openserv.ai The agentic economy won’t scale on better models alone. It needs better reasoning infrastructure that are finetuned to the reliability, auditability, and affordability requirements of enterprise agents. SERV is building that layer.

Introducing SkyTail, and building XBox games with Three.js and @Geastack SkyTail is our very first game that we developed at @coyotiv. Besides how fun it is to play, the most striking feature is that it's a JavaScript game built with Three.js, compiled natively to Mac, iPhone, iPad, Windows, and most importantly, XBox, to run at full native 4K resolution at 60 frames per second. Gea Stack allows you to build native applications for embedded devices, computers, phones, and game consoles with the same familiar developer experience you know and love from the web. piped.video/watch?v=ETfSiUHe…

I got access to Jev earlier today (thank you @hackgoofer). I have run ~5,000 requests so far, (which cost me around $2!), across classification, model routing, intent, steering, and many other things. tl;dr, Jev enables a new intelligent decision-making primitive, separate from deterministic code and LLM calls. This allows a class of decision-making that was neither suited to dumb, unintelligent code, nor to slow, expensive LLMs. It is super fast and cheap, and I think I will likely end up making a few Jev calls to every LLM call I make in my product. I think probably any company using LLM requests today can probably add a Jev call pre and/or post LLM calls to quite literally make their product much better for free, and have better tool calling behavior in many cases. I happened to have a personal benchmark for this as I’d been working on a ton of proactivity and classification tasks. I have been using the deepseek flash and more recently gpt 5.6 luna family of models as reasonably smart classifiers with low latency. Think questions like: - Did this conversation output contradict something they’ve mentioned before? - Should we send a followup message to this user based on our rules? - It’s been a few seconds of silence. Should we proactively send a message? In the past, I’ve been forced to write a bunch of what I call decision chains, mostly because an LLM classification call is very expensive in TIME (avg 4s), and less importantly can cost quite a bit if run on every message. Imagine a normal chat app. If you added 4s to every response to figure out if the response is good before sending it out, that ends up being pretty bad. So instead, I usually have to write some code that is a crude heuristic that runs quickly and decides whether to run the classifier. Obviously this sucks because you call the classifier many times that you don’t want to, which makes your p95 bad, and you also miss cases with the heuristic, and you also have to manage all of these weird chains. With Jev, it’s cheap enough, and fast enough (p50 ~150ms, p95 ~350ms in my testing!) that you can easily run it every turn. Heck you can reasonably run it before generation AND post generation, for any application that isn’t realtime voice, and still feel snappy. But this is just one use case. Think: smarter model routing, better context packing, better responses, better observability for intent/tags/safety, smarter retries and so much more. By simply thinking about the inputs and outcomes you want to enforce, you can use Jev to supercharge most model calls and reduce bad user outcomes. The more “quirks” a model has, the more valuable it ends up being. It’s a bit weird and unintuitive using Jev. Generally, you want to decrease the # of questions you ask a classifier, or it makes more mistakes. In fact, you might want to ask your questions kind of in a compound way, because the reasoning happens in a shared scratchpad of sorts. Adding questions muddies the scratchpad and makes it take longer. With Jev, you feel incentivized to go the other way, to formulate your query as a set of independent questions. It doesn’t feel like adding more questions decreases your performance on others. You can go a bit deeper to improve tool calls. Many model tools are things like turning on settings, or other things. You can easily improve models that are not very good at tool calling with Jev, by simply figuring out when to run them. You can do a pre-LLM call to figure out when to unfurl different tool definitions, in order to make your main LLM run better, you could run a background task with Jev + another LLM to reduce tool and context burden on your main LLM, and free it to be responsive. I’ve only scratched the surface of my testing but very excited!

12M+ views so I had to dig into this Is it just a hyped launch or something novel and actually useful? I’m leaning to the latter and will break it down with a practical example First, understand that Jev isn’t a typical LLM that you prompt and get back a written response It’s much more specific - you give Jev context along with a set of questions and possible answers and it returns probabilities to each question and answer set, and it does it super fast and cheap I found it helpful to think through a simple example: Let’s say you’re an online retailer and you get hundreds of inquiries a day You put an agent in the workflow that looks at each inquiry and makes a judgement call on the next action and then generates a response But that LLM judgement call can be messy - LLMs still have a tendency to make shit up or respond in ways that you don’t like, so we put humans in the loop What Jev does is look at each inquiry and makes a fast decision on the nature of it so the workflow can take the next best action with confidence It doesn’t replace the workflow or the LLM entirely, it handles the fuzzy fork in the road that used to require a human to read the message first Let’s say you get a customer email: “This jacket sucks! The zipper jammed the first time I wore it and now it won’t close so I want my money back.” That message lands in the retailer’s system along with a few facts the company already has: the order is 11 days old, the return window is 30 days, the jacket was $100, and this customer has one previous order with no refunds In a typical LLM workflow you’d let the model decide what to do and have it initiate that action and send a reply email, BUT many companies want a HITL in many of these cases, especially ones where it’s not entirely clear what to do Imagine the LLM replies to the email with troubleshooting instructions on how to fix the zipper - probably not the best action With Jev, the company doesn’t ask it to write a reply, it asks a few specific questions about the email and order >Does the customer want a refund? >What is the main issue: defective product, didn’t like it, late delivery, or something else? >How unhappy does the person sound? >Does this look like it falls inside the published return policy? >Should this go to returns, quality control, or a human agent? Jev looks at the email and the order context, then answers all of those at the same time with a probability score - no prose, just a score In this case the picture comes back fairly clear: yes, this is a refund request, the issue is a defective product, frustration is high but not explosive, the request is inside the return window, returns should own it That is enough for the retailer’s existing rules to take the next action with confidence If Jev had been less sure - say the email was sarcastic, the order was 45 days old, or it was unclear whether the customer wanted a refund or a replacement - the workflow could put the ticket in a human queue instead of guessing The useful part is the handoff, specifically knowing when to pass it to the next step in a workflow or route to HITL If Claude were the judge in a less clear situation, you’re likely to get a long winded answer that still leaves you wondering what to do next, and it would have taken longer and cost a lot more This was a simple example but I can see how this can be incredibly useful in complex enterprise workflows which is why I lean towards novel and useful over just a hyped new thing Will be following this one closely

The 1st SERV Hackathon is live. Two weeks to build with SERV Reasoning and get rewarded. Get hands-on early before SERV v3/v4 brings graph sharding, self-improving intelligence, and task-specific models. Submissions close Sept 28. Pick your track: openserv.ai/hackathon

Everything we’ve done so far was preparation for this. Today, we open access to production-grade AI reasoning for every team. SERV Reasoning API is now live. This unlocks next stages: SERV v3→v4. Make your first API call & check 26/27 milestones: openserv.ai

Everything we’ve done so far was preparation for this. Today, we open access to production-grade AI reasoning for every team. SERV Reasoning API is now live. This unlocks next stages: SERV v3→v4. Make your first API call & check 26/27 milestones: openserv.ai