
I test AI tools in real-world use and share what actually works for me. I also talk about technology and other topics that interest me. @Hitvox
Brasil
Joined August 2009
- Tweets1.7K
- Following4.6K
- Followers663
- Likes17.1K
Liquid AI released d1.2 and d1.2-ultra, decision models that now support audio input and five new task types.
They can extract values from documents, find every relevant passage, locate objects in images, estimate quantities, and select all applicable options. Unlike generative models, they return structured decisions with calibrated probabilities.
Today, we release d1.2 and d1.2-ultra, the state-of-the-art omnimodal decision models. d1.2 and d1.2-ultra both can process text, images, and audio with new output capabilities.
> #1 on the @ArtificialAnlys Decision Index (75)
> Read a value from the input (text, number or date)
> Find every passage of a kind in a text (spans)
> Locate things in a picture (box for one, boxes for every instance)
> Estimate a number (number with a range)
> Pick every option that applies (choice)
1/ 🧵
Tencent introduced Hy4 flash, a multimodal model designed for everyday productivity at a much lower cost.
It supports a 1M-token context window and is priced at $0.07 per million input tokens and $0.281 per million output tokens. Tencent says its performance in daily work is close to Claude Haiku 5.5 in blind evaluations.
⚡Meet Hy4 flash, the cheapest model for everyday productivity. Matches Claude Haiku 5.5 (high) on daily work (human blind tests), yet 18x task cost reduction. API pricing (per 1M tokens): $0.003 cached · $0.070 input · $0.281 output. 🔗:hy.tencent.ai/research/hy4-f…
29
Qwen has introduced Qwen4-Max, a new AI model with native vision, a 1M-token context window, and a 2.4T-parameter backbone.
It is designed for long-horizon work across coding, 3D modeling, game development, manufacturing, finance, law, healthcare, and scientific research.
📢 Meet Qwen4-Max, your new AI teammate for work, creativity, and discovery!
Qwen4-Max is built on a new MoE architecture with native vision and a 1M-token context window.
Highlights:
- Outperforms Qwen3.8-Max with up to ~50% fewer tokens, while average cost per task drops over 60% in our internal evaluation.
- Advances real-world productivity through greater proficiency with professional tools, stronger long-horizon execution, and higher-quality deliverables.
- Handles real tasks, including 3D modeling, game development, manufacturing, finance, law, healthcare, and frontier scientific discovery.
- $1.5/M tokens for input, $4.5/M tokens for output, and $0.075/M tokens for implicit caching — all lower than Qwen3.8-Max, with implicit caching down by 70%.
The API is live on QwenCloud. You can also try Qwen4-Max on Qwen Studio, QwenWork and Qoder.
Start building today! 🚀
- Blog: qwen.ai/blog?id=qwen4-max
- API: qwencloud.com/models/qwen4-m…
- Token Plan: qwencloud.com/pricing/token-…
More to come! The open weights of smaller Qwen4 models will be released. Stay tuned. 👀
38
Daivison Cardoso retweeted
uma boa documentação de design é essencial pra conseguir dar mais liberdade para os agentes construírem a feature de cabo a rabo
eu praticamente não preciso mais instruir o agente com ajustes finos de UI, minhas iterações com design são quase todas voltadas a fluxo do usuário
Daivison Cardoso retweeted
I’ve compiled a list of all 705 sandbox companies
orble.tech/sandboxes
Daivison Cardoso retweeted
update on my open-source RSI coding agent...
Claude and Codex now run inside DeepSeek Harness entirely through T3 Code.
and Google's Harness² is hooked up, so the agent can test changes to its own prompts, skills and tools to improve itself.
DSH is fully in charge of both of them too:
> it sets their modes, instructions and tool permissions
> they run my DSH plugins directly, with no coordinator model in between anymore
> their threads stick around, so follow-ups keep the same context
next I'm tying it all into one loop: Claude builds, tests run on their own, Codex reviews and fixes the work, and Harness² keeps improving the whole system.
I'm trying to build my own open-source RSI coding agent...
yesterday I had the idea, so I immediately installed T3 Code and DeepSeek Harness and started wiring them together
by last night, Codex inside DSH was handing work to Claude through T3, all on my existing subscriptions
Theo, please don't hate me for using T3 Code this way lol
T3 Code is built as a frontend, so ironically it works perfectly as my backend, now every session shows up in a super clean UI I can browse and search in seconds
that first version still used a coordinator model to pass tasks along, so now I'm building a direct bridge, a custom DSH plugin that talks to T3 through MCP, so I'm not burning extra tokens on a model just to route tasks
next up I have some fun ideas:
> DeepSeek's Super-agent design to split work across parallel workers
> Google's Harness² to rewrite the agent's own prompts, skills and tools overnight and keep only the changes that help
once the loop is live, I'll post the full setup and open-source the whole thing if you guys want it
People keep asking how I actually use Tucano day to day.
This video is one of my workspaces: the team that helps me build Tucano itself. 14 agents in 3 sectors, and I only talk to one of them.
I'm the only maintainer. There's the desktop app, iPhone, Android, PR review, store releases, customer email, posts, Discord. That doesn't fit in one person. It doesn't fit in one agent either: ever tried setting up a single Claude or Hermes to handle all of it? You end up with a giant prompt that does everything halfway.
In Tucano I split it up like a company:
- Orion is my entry point. I ask him.
- Engineering: Bento leads, Vera reviews, Nilo handles mobile, Otto ships.
- Customers: Clara and Caio.
- Growth: Iris, with one agent per network.
In the video I send a single sentence: "Caio says an attached PDF only arrives as text, fix it and let everyone know."
Orion passes it to Engineering and Growth. Bento splits the work. Engineering asks Customers to draft the reply to the person who reported it. Every line that lights up is a request with an owner, and the answer travels back the same way until it reaches me.
Two rules hold it together:
- An agent never talks to an agent in another sector. Only head to head.
- Nothing goes out without my OK.
The bug in the video was real: @caiotferreira reported it today and the fix shipped in 0.2.31. The video is a demo of how a request like that moves through the team.
Tucano is open source and runs on your own computer, on the Claude Code or Codex subscription you already pay for. Download it, create two agents, connect one to the other and send a request. Watching that first line light up is what sells it.
Daivison Cardoso retweeted
MyGo : un framework pour construire vos apps desktop macOS, Windows et Linux en Go, avec une UI native dessinée sur le GPU ou un frontend web dans la webview du système. ⬇️
Google introduced Gemini 4 Hydrogen, a faster model for everyday coding and agentic workflows.
It supports four thinking levels, a 1M-token context window, and can work independently or alongside Gemini 4 Carbon and Argon. Google says Hydrogen completes tasks 2.5× faster than Gemini 3.8 Flash while improving coding performance.
Meet Gemini 4 Hydrogen.
It brings cutting-edge intelligence to everyone, with speed and efficiency for everyday coding and agentic work.
Across 4 thinking levels, it stands on its own or teams up with Carbon and Argon — 2.5X faster than 3.8 Flash, outpacing competing models in its class.
1
1
60
Tucano 0.2.28 🧡
You asked for it, a lot: agents now have an org chart.
Talk to one. It splits the work across the team and closes the subject with you.
Pull a line from one agent to another and watch the request move.
Also: Hermes as a harness, Codex limits, duplicate routines.
Release Notes github.com/eduwxyz/tucano/re…
Google’s Gemini Decisions API is built for fast, structured decisions instead of generating long responses.
Gemini Decision 1 can evaluate text, images, audio, video, and PDFs against up to 1,000 questions in a single request, returning probabilities and confidence scores. It costs $0.10 per million input tokens, with no output-token charges.
We're introducing the Gemini Decisions API in @GoogleAIStudio, powered by our new Gemini Decision 1 model.
Instead of generating a text response token by token, Gemini Decision 1 evaluates your input (across text, images, audio, video, and PDFs) against up to 1,000 questions to make fast, confident decisions in a single pass, with zero output tokens billed.
Here's a look at how it works 🧵
16
Gemini 4 Carbon is now available as Google’s strongest model for long-horizon agentic workflows, complex software engineering, cybersecurity, and scientific reasoning.
While Argon is aimed at most coding and knowledge work, Carbon is designed for harder tasks that require deeper reasoning, extended verification, and multiple steps.
Gemini 4 Carbon is now available.
It’s our strongest model, built to dig deeper, handle long-horizon tasks, and multi-step reasoning.
If Argon is great for complex tasks, Carbon is the model you bring in for your hardest problems.
25
MiniMax has released H3-Image as an open-weight model for image generation and editing.
It supports up to 16 reference images and can generate images of around 16 megapixels, including 4K outputs.
4K image generation in 10 seconds! ⚡
Meet MiniMax H3-Image.
16 million pixel resolution. Stunning realism. Precise prompt following & editing. Up to 16 reference images. All in 10 seconds.
And it's open-weight. 🚀
Available now via our API, MiniMax Code, and Hailuo AI — with 50% off 16MP+ generations during launch week!
Can't wait to see what you create. 🔥
Hugging Face: huggingface.co/MiniMaxAI/Min…
API: platform.minimax.io/docs/gui…
MiniMax Code: code.minimax.io/about
1
51
Daivison Cardoso retweeted
Pra quem não sabe, eu tô começando uma nova startup, a Tesar (tesar.ai) 🛰️
A ideia: monitorar infraestrutura física (linhas de transmissão, dutos, barragens, minas, industria) com satélite, clima e terreno, câmeras, drones e sensores no geral. Tudo isso visando suportar operadores e industrias!
O desafio técnico é absurdo e eu vou começar a contar aqui como a gente tá construindo isso tudo.
Build in public começa!
Meta introduced Muse Ember, its smallest and most efficient model for reasoning execution and repeatable agent work.
It can call tools sequentially, run coding tasks, parse long documents, and handle workflows across apps. With a 1M-token context window, it supports text, images, audio, and video in the same workflow.
Today we're introducing Muse Ember, our smallest and most efficient model, built for complex reasoning execution and everyday agent work.
Muse Ember can carry workflows across apps, call tools in sequence, run tests and make targeted code fixes. With a 1M-token context window and native understanding of text, images, audio and video, it can also work across large document sets and varied inputs.
2
110
Anthropic introduced Claude Fable 5.2 and Claude Mythos 5.2, with improvements in coding, reasoning, and vision.
During testing, the models supported complex software migrations, robotics evaluations, hardware design, and research in physics and mathematics. They are also reported to generate outputs up to 30% faster than the previous versions.
Introducing Claude Fable 5.2 and Claude Mythos 5.2, designed to tackle the most daunting technical challenges.
They’re capable collaborators on ambitious projects in many fields, and show improvements in areas like coding, reasoning, and vision.
81
Daivison Cardoso retweeted
Releasing Muse Ember today -- our smallest model that delivers higher intelligence per dollar than others in its class. Built for token-heavy work that agents repeat all day: tool calls, coding subtasks, routing, extraction, and long document parsing. Available today on the Meta Model API and in Muse Code.
Daivison Cardoso retweeted
Want your own Slack for $0? I open-sourced Ember. It deploys to @Cloudflare’s free tier in about a minute, works with Slack-compatible payloads, and includes an AI assistant and media attachments.





























