Today we are open-sourcing Matilda-K3, a 2.8T-parameter frontier model post-trained in Australia with our own conditional post-training. Conditional post-training changes how a model behaves on the requests we target and leaves every other request to the frozen base model, so the changes are precise and the original capability stays intact. Matilda-K3 is built on Kimi K3, the open-weight model from Kimi (Moonshot AI). The strongest open-weight models today come from a small number of labs, and a model carries the assumptions of the lab that trained it. Our post-training keeps the base model's capability and changes how it behaves in two targeted areas: its identity and its biases. The base weights are frozen and shipped unmodified. Everything else is left unchanged. It holds its identity under pressure, presenting as Matilda in all 474 held-out adversarial prompts, including jailbreaks, role-play and long-document attacks. On contested political questions, balanced or facts-only answers rose from 9% to 92% and one-sided answers fell from 54% to 3%, with factual accuracy unchanged. It keeps the full capability of its base: 95.0% on AIME 2025 and 99.4% on HumanEval, both within run-to-run variation of the original model, with a 1M-token context window, native reasoning, and image and video input. This approach can be applied to shape how a frontier model behaves for a specific organisation, including identity, tone, and policy, without sacrificing its capability.

Today we are open-sourcing Matilda Jev, Australia's first decision model, conditioned for Australian deployment and post-trained with our own specialised recipe. Decision models do not generate text. They answer a typed question and respond with a calibrated probability for each option in a single forward pass, reducing hallucination and making intelligent decisions faster and more efficient. This release proves that highly capable AI models can be built, post-trained, evaluated, and served entirely within Australia. Post-trained end-to-end on our local infrastructure, the model achieved a median decision latency of 56.8ms on a single @AMD MI355X across 150,317 evaluation requests.

Almost everything genuinely known about serving large models fast lives inside a handful of companies, and almost none of it gets written down. That's why we decided to put all our learnings into a book that you can have for free. From Tin to Tokens is a free 192-page book on building an LLM inference stack from scratch in Rust. It goes from the driver interface, where you are writing ioctl structures by hand, to an HTTP endpoint serving thousands of concurrent users: attention kernels, a paged KV cache, a scheduler, multi-GPU collectives, a quantisation pipeline, and the correctness apparatus that proves any of it works. It came out of the work we do at Maincode, where we build frontier AI in Australia on our own machines. The inference layer is where the physics of the hardware meets the economics of the product, and it is where most of the difference between a demo and a business gets decided. And we made it free because this kind of knowledge needs to be shared, and it's super cool to boot! And if you find something wrong, which I'm sure you will, let us know! Link to the repo is in the comments.

We wrote our own inference engine in Rust 🦀 for our AMD MI355Xs, with no ROCm anywhere underneath it. On the small all-reduce that tensor-parallel inference runs for every single token, it comes out about 1.3x faster than AMD's own RCCL at its best, sitting on what looks like the physical latency floor of the chip at around twenty microseconds. As the first anywhere to have the MI355X fully deployed in production, running Matilda, inference is a huge focus for @MaincodeAU. In fact inference is the part of an AI business that actually meets customers, so we wanted to understand this hardware as deeply as we could. We took the runtime out and drove the GPU through the Linux kernel driver directly, the same interface ROCm itself is built on, to see how close to the silicon we could get. RCCL is genuinely great and has years of careful work behind it, and it also has to be good at everything a runtime is ever asked to do. A layer that only has to do one thing really well can set all of that aside and cut its overhead to almost nothing. The win isn't a clever trick, just the absence of overhead. And you don't have to take our word for it. The whole thing ships as a single container image with the kernels compiled in, and three commands reproduce the numbers on any MI355X box. More in the blog post...