agua

Hello, I'm awa, creator of OpceanAI research and OpceanAI lab

[agua @awa_omg]

Moud — Public Beta

After a few days of development, testing, breaking things, fixing them, and probably breaking them again, Moud is finally entering Public Beta.

What started as a fairly simple idea grew into something much bigger, much faster than we expected.

Moud is an AI and developer infrastructure platform built by OpceanAI, bringing models, APIs, agents, compute, deployments, and developer tools together in one place.

289 models

Moud currently gives you access to 289 models through a single platform.

The catalog includes language and reasoning models, along with other AI capabilities provided through different providers.

We also host a selection of models ourselves. These are part of Moud's own infrastructure and are available under the conditions of the relevant service.

The catalog will keep changing as models, providers, and infrastructure availability change.

More than a model playground

Moud isn't just a place to try different models.

You can also build on top of them with:

- API access
- BYOK
- Agents
- Compute
- Virtual machines
- Deployments
- Managed DNS
- Moud subdomains
- Channels and integrations
- Developer tools

You can also register a device and sign in without having to copy and paste an API key every time.

Privacy

Moud is built with privacy and data minimization in mind.

Selected services hosted directly by Moud operate with Zero Data Retention (ZDR).

Third-party models remain subject to the privacy, retention, and data-processing policies of their respective providers.

Our infrastructure operates across Mexico, Canada, and the United States, with Moud's primary operations based in Mexico.

So... how are we paying for all of this?

Well...

One person's money is another person's tokens. :)

Seriously though, Moud relies on a mix of our own infrastructure, provider capacity, shared resources, and usage limits to keep a significant part of the platform available for free.

And yes.

All of this happened in a matter of days.

So this beta should be interesting.

There will be bugs.

Models may disappear.

Providers will probably have problems at some point.

Things will change.

And we'll probably break something ourselves every now and then.

That's exactly why we're opening the beta.

We want people to actually use Moud, find the rough edges, tell us what's missing, and help us figure out what to build next.

So go ahead and burn our money.

The bank is definitely going to kill us when it sees the statement in three days, but that's fine. :)

Moud Public Beta is now open.

Welcome to Moud.

https://t.co/gFi4vI3QKt

[agua @awa_omg]

@AnthropicAI @claudeai For context, the Free plan only includes access to Sonnet and Haiku. Access to paid-tier models should require the appropriate subscription.

[agua @awa_omg]

We are investigating a potential vulnerability in Anthropic’s subscription enforcement that may allow continued access to paid models despite an inactive subscription warning.

We are documenting our findings for responsible disclosure.

@AnthropicAI @claudeai

[agua @awa_omg]

@claudeai Thanks for the clarity. I claimed our Claude Team year on Oct 7 (invoice shows the promo applied), but it was cancelled after a mistaken suspension that Safeguards reversed on Oct 9. Since claimed offers remain, could you help restore it? Support ticket is open.

[agua @awa_omg]

@duolingoespano @duolingo que haces aquí fred? 😭 hasta en la escuela me sigues

[agua @awa_omg]

ok so I got accepted into the Claude Startups program

btw one day later they revoked my account and took the benefits away 🥀

first time ever using Claude, building . appeal is in, ref TS-01a11bd9-. would love a human look @claudeai @AnthropicAI

[agua @awa_omg]

@rodydavis @AndroidDev @Android @antigravity It looks like there are still a few details that need to be fixed.

[agua @awa_omg]

@Badtheorylabs We appreciate you too ❤️ Glad we could help. Best of luck with Tinfield 1!

[agua @awa_omg]

Today, OpceanAI is supporting a startup for the first time.

We’re providing @Badtheorylabs with compute capacity to help bring Tinfield 1 to production.

Their API is now live at https://t.co/nW9qxpPZJH.

Glad to support the team.

[agua @awa_omg]

@Badtheorylabs Hi! We saw you need compute to deploy your model, and we’d like to help.

We can provide a 2× RTX 5090, 193 GB RAM, AMD EPYC 9384X (32C/64T) for inference.

It may not cover all your needs, but could help with the initial load while you secure more infrastructure.

Interested? DM

[agua @awa_omg]

@dphnAI You can use OASIS through our public OpenAI-compatible API:
https://t.co/cewrDGglqm
We also have our own chat at https://t.co/LQX3Pzkvve, with audio and image support on compatible models, plus https://t.co/iYV5VOKHQ2 for infrastructure monitoring.

[agua @awa_omg]

@hatsutahaku

[agua @awa_omg]

Introducing PASITA v1.

An 88M-parameter language model trained entirely from scratch for one task: converting plain text into valid Markdown while preserving the information in the source.

Our goal is simple: add structure without adding content.

We think of PASITA as compiler behavior rather than chatbot behavior.

On our private held-out PASITA-bench-1000 evaluation:

• 95.6% GFM validity
• 92.7% faithfulness
• 88.8% semantic faithfulness
• 100% table fidelity

PASITA is designed for document-processing workflows including OCR post-processing, document structuring, table preservation, HTML extraction and RAG preprocessing.

The model uses a dense "LlamaForCausalLM" architecture with 12 layers, GQA, a 16K byte-level BPE tokenizer, and a 2K context window. It supports Spanish and English.

Training combines SFT, DPO and GRPO, with verifiable rewards targeting formatting validity, numeric fidelity and anti-overformatting.

PASITA is intentionally specialized. It is not a general-purpose instruction-following model, and we document its current failure modes and limitations in the Model Card.

Apache-2.0. Open weights.

Model and full technical details:

"https://t.co/bgMOJTH0Nq"

A small model for a constrained problem.

PASITA v1 is our first step toward exploring how far specialization can take us.

[agua @awa_omg]

Hello, Moud community.

We’ve been relatively quiet over the past few days, and there was a reason for it. We’ve been working on a surprise.

Sunlight is now available.

Sunlight is an Android app designed to run AI locally, use models through BYOK, and, of course, access Moud’s AI.

The current release includes:

1. Web Fetch
2. Document creation
3. Local AI inference
4. Image vision and audio detection

Sunlight is built for Android ARM64 and supports additional security through fingerprint or Face ID authentication, as well as 6-digit code-based login through the Moud dashboard.

Here is an example of document creation with Sunlight:

https://t.co/YhTJyr9u6z

This is just the beginning. More capabilities are coming to Sunlight.

https://t.co/orultlYX6g

[agua @awa_omg]

@Hostinger Claude Code is really taking off, huh?

By completing your purchase, you agree to our <a href="Terms of Service" target="_blank" data-qa="purchase-details-terms-of-service"> <strong> Terms of Service</strong></a> and confirm that you have read our <a href=" Privacy Policy" target="_blank" data-qa="purchase-details-privacy-policy"> <strong>Privacy Policy</strong></a>.

[agua @awa_omg]

Guys, Ox-Alpha gave us quite a lot to talk about 👀

Error: 400: {"type":"server_error","message":"Error from provider (Console Go): Upstream request failed: [1210] 图片输入 格式/解析错误"}

What’s interesting is not just that the error is in Chinese.

The [1210] code, together with the message “图片输入格式/解析错误” (“image input format/parsing error”), matches errors documented in Zhipu AI / GLM infrastructure, particularly when processing visual inputs.

[agua @awa_omg]

Just to clarify, the following models will remain available without any changes. The complete list is therefore:

- LFM2.5 1.2B Thinking
- LFM2.5 2.6B
- LFM2.5 8B
- Qwen3.5 2B
- Qwen3.5 4B
- Ling Tiny
- Gemma 4 2B
- Ornith 1.5 9B
- Gemma 4 4B

[agua @awa_omg]

Moud — Hosted Models and Public Beta

We’ve decided to take another step with Moud.

Alongside the models already available on the platform, we’re now starting to host a selection of smaller and specialized models ourselves.

They’ll be free to use, and their usage won’t count against your regular Moud token limits.

The idea is simple: keep a set of useful, lightweight models available for everyday tasks without requiring access to a large model every time.

Small Models

- LFM2.5 1.2B Thinking
- LFM2.5 2.6B
- LFM2.5 8B
- Qwen3.5 2B
- Qwen3.5 4B
- Ling Tiny
- Gemma 4 2B

Medium Models — approximately 4.5B–9B

- Ornith-1.5-9B
- GLM-4.6V-Flash — 9B Vision
- Qwen3.5-9B-Base
- Gemma-4-E4B

Specialized Models

- GLM-OCR
- PaddleOCR

These are hosted directly by the Moud team. We’re running the inference infrastructure ourselves rather than simply forwarding requests to an external provider.

This is also part of where we want Moud to go: one place where you can access models and AI services, use APIs, run agents, deploy projects, and eventually get the compute you need to build things with them.

We’re also working on Moonlight, our own AI agent and developer CLI. We’ll have more to share about that soon.

As for the Public Beta, we’re currently aiming for a release somewhere between Sunday and Wednesday.

We’re getting close now.

More soon.

[agua @awa_omg]

Guys, a quick update: due to current computing capacity constraints, we regret to inform you that we can currently only support the larger Ornith and Gemma models. We hope to expand our capacity and support additional models soon. :(

[agua @awa_omg]

@404malwarekitty siento que una animación mientras los comandos aparecen y muestran el contenido aportaría más una fluidez visual, también la vista móvil se ve muy apilada y hace que la lectura se dificulte un poco