When are there going to be price/duration pricing for LLMs? Mercury 3 is running at 1000 tok/s $0.70 per million output 16.667 min to one million output tokens (uninterrupted token streaming) Vs Haiku 5.5 runs at 232 tok/s (according to @ArtificialAnlys) $0.50 per million output tokens ~72 min to one million output tokens So you’re [let’s call it] investing your $ into tokens, 5x faster (omitting price). Of course, this isn’t a big deal because you would ideally get to your outcome faster and still spend the same amount of tokens (if not less because Mercury 3 is reportedly smarter than Haiku 5.5).

Actual “decision tree”. @Raspberry_Pi 5 running @liquidai d1 600m Being powered by solar With a diy tree connected to Figuring out how to productize this for < $125

Had a good time going to the Open by Design event hosted by @reflection_ai @nvidiaai @togethercompute with @electricaliee If you’re interested in local, edge, open source ai, let’s chat!

Try out @liquidai new LFM 2.5 24B model routed by their D1 decision model, made by us @offlinebase. Website in comments. @EthanCastr

Today we are announcing our biggest bet yet on our own silicon. Starting next calendar year, we will scale our Maia family to multi-gigawatt deployments, making it the primary inference system for our own product and services. And we’re introducing Maia 300, which delivers 40% more tokens per watt compared to leading GPU-based systems. We are already deploying Maia 200 across the US, Europe, and Australia, delivering up to 1.7x more tokens per watt across experiences in Teams and Copilot. Our goal is to build on the full-stack efficiency gains we’ve already seen, while continuing to bring the latest innovations from across the ecosystem into our fleet.

Qwen3.8-27B at #9 overall on Code Arena, the only model in its size class in the top 10.😎 Small but mighty. Thanks for the recognition! @arena

Had a great time at @cerebras cafe compute with @habodoni @electricaliee and @Norminican at the @datadoghq office Thank you to the hosts @zhennydez @MilksandMatcha @thorwebdev and more.

Here's one of our (@Habodoni @Electricaliee and I) submission to the @cerebras @googlegemma Project Description: An app for people prone to injuries or accidents. Use cerebras to react in a split second to emergencies. Sub agents spawn and call/email/text 100+ people your location, a description of your setting, resources nearby, and more within 2 seconds. Likened to a faster, smarter, better, multimodal, agentic Life Alert!

Here's one of our (@Habodoni @Electricaliee and I) submission to the @cerebras @googlegemma Project Description: Upload photos of ANYTHING (IDE, new software, gym, fridge, construction materials) and get a walkthrough, image labeling, deep research on specific things in the photo, and share your creations via a link! Everything is done within 3 seconds across up to 15 sub agents

Here's one of our (@Habodoni @Electricaliee and I) submission to the @cerebras @googlegemma hackathon. Project Description: Ferry is a local-first AI gateway for unreliable internet. Open WebUI talks to Ferry like a normal OpenAI-compatible model endpoint; simple prompts are answered instantly by a local on-device model, while harder prompts are stored in a local SQLite backlog queue. When connectivity returns, Ferry bursts the queued work to Gemma 4 on Cerebras and streams the answers back into the same chat experience. All you need is 5 seconds of internet connection to get all of your cloud AI requests that couldn't be answered locally.

@GoogleAI @Google @googledocs @GoogleWorkspace Can y'all add a bit. ly type link shortener and QR code generator to google slides?

I also think @Replit , @vercel v0, @Lovable and the rest should consider using @cerebras and @GroqInc for MVPs and prototypes. It's very hard to edit and augment when you have to wait 2-3 min for the agent and it may/may not do what you want. Also, consider the price of something like Kimi or GLM or Qwen, it is on average 1/3 of the price of a frontier company's LLM and you can still have Sonnet/Opus or GPT-5.1 be the orchestrator. @Replit maybe you should add a fast design mode the same way you've added the fast mode for quick edits. Of course, if you want to just retain your partnership with Anthropic (and maybe google with gemini 3), just use Haiku 4.5 or Flash 2.5:preview. Theoretically if a code editor uses cerebras or groq, they may be able to create 3 previews of a site within a few seconds, for someone to do a visual choose your own adventure. @amasad has discussed how PMs are expected to come up with a mock or demo to hand off/show engineers. It's not so great if you start off with mid, then iterate upon mid. Also, we are not yet at the point where everyone can accurately externalize their thinking via a prompt, but we are closer to folks being shown X options and them selecting which one is closest to what they were imagining. @mattyp, @amasad, @rauchg, @SarahChieng, @ozenhati, @JonathanRoss321 thank you all for having a public presence and making me aware of these tools/your tech. @chamath your mention of @GroqInc , last year is what introduced me to the world of inference, so thanks.

Working on a small project to demonstrate the value of faster inference augmenting.work/ Thesis: Frontier models have pretty fast inference, though not fast enough for test-time compute/reasoning and 10+ tool calls. If we are to augment with AI and save time, let's not create a new type of down time... waiting for responses. We used Cerebras as our inference provider to create websites, analysis, and textbook chapters within 2 seconds. We typically hit around 2000 tokens/s. Details: Using Cerebras and Groq as inference providers. Models - GLM 4.6, GPT-OSS-120B, Qwen... Horrible demo video below. Only the typing is sped up. I am still working on this. I just wanted to start posting updates for posterity because @GroqInc and @IBM partnered after I started this and I don't want to miss anymore news regarding the benefit of fast inference when it comes to tool-calling, Reasoning, Enterprise, and healthcare need for speed.