Introducing Homan 2.96: Our Most Capable Model Yet

Today marks a milestone for Bloom Labs and, we believe, for the entire field of machine intelligence: we have completed internal development of Homan 2.96, our new flagship model and the single most capable system we have ever built. This isn't an incremental update. It's a ground-up leap in reasoning depth, coding ability, multimodal understanding, and working memory — and our team has spent weeks putting it through its paces internally.

We built Homan 2.96 to answer one question: what happens when you stop optimizing a model for benchmarks and start optimizing it for genuinely difficult, open-ended work? The result is a model that doesn't just answer questions — it reasons through them, plans multi-step solutions, writes and debugs production-grade code, reads images and diagrams with real comprehension, and holds an entire codebase or research paper in its head without losing the thread.

Headline Capabilities

• Deep, deliberate reasoning. Homan 2.96 can work through multi-step logic problems, mathematical proofs, and ambiguous real-world questions with a level of care that feels less like autocomplete and more like thinking out loud — because, in a very real sense, that's exactly what it's doing.
• State-of-the-art coding. From scaffolding an entire application to hunting down a single off-by-one bug buried in a legacy codebase, Homan 2.96 writes, explains, and refactors code across dozens of languages with a fluency that outpaces every prior Bloom Labs model by a wide margin.
• True multimodal understanding. Homan 2.96 doesn't just accept images — it reasons about them. Charts, screenshots, hand-drawn diagrams, photographs of whiteboards: all become first-class inputs it can analyze, describe, and act on.
• Massive long-context memory. Feed it an entire technical manual, a sprawling codebase, or a stack of research papers, and Homan 2.96 keeps track of it all, cross-referencing details from the very first page while reasoning about the last.

Why This Is a Leap, Not a Step

Every prior Bloom Labs model has been a meaningful improvement over the last. Homan 2.96 is something different. Internally, we've been stunned by how often it produces answers our own research team wouldn't have thought to check for — catching edge cases in code, spotting flawed assumptions in a question, or offering a more elegant approach than the one we asked for. It is, quite simply, the smartest thing Bloom Labs has ever built, and we think it will change what people expect from an AI model once the world gets to see it.

This milestone also sets the tone for our full model lineup: Homan 2.96 sits at the top as our flagship, with Bloom 4.6 delivering that same DNA in a balanced, everyday package, and BabyMax 5.76 bringing it to lightweight, blazing-fast use cases. Together, they represent the most advanced family of models Bloom Labs has ever built — and, for now, the most advanced family we're keeping in-house.

Homan 2.96 has completed internal development and testing, but it has not yet been released publicly and is not accessible to anyone outside Bloom Labs. Given how capable it is, we are being extremely careful: our safety team is conducting a full risk review, and we will not ship it until we are confident it cannot cause harm. We know that's not the announcement people were hoping for today, but we'd rather take the time to get it right.

Comments (27)