FRAMERAI

What’s New

1. Aug 27, 2026MilestoneLatest

   FramerAI 3T: three trillion parameters, one million tokens of context↗

   • framer-3t-a64b is the new flagship: 3.01T parameters total, 64B active per text token.
   • 2.93T in the backbone, 76.7B across the vision, audio, image, video and audio-LM towers.
   • Towers nearly doubled against the 2T preset, so image, video and audio scale with the backbone instead of trailing it.
   • 512 fine-grained experts at top-6 routing, dividing evenly across every expert-parallel mesh width.
   • 1,048,576-token context on all three trillion-scale presets, via YaRN RoPE scaling.
   • Context costs no parameters and 336 GiB of KV cache per sequence, which --estimate reports.
   • Validation rejects a context no scaling factor covers; prefill runs in bounded chunks so the window is reachable rather than nominal.
   • Documentation, band tests and 743 passing tests shipped with it.
2. Aug 15, 2026Milestone

   The cognition layer↗

   A trained model answers every question from what it learned during training plus whatever is in front of it right now. When the conversation ends, it keeps nothing. This change added an optional layer that gives it a history.

   It is entirely opt-in. Nothing else in the system depends on it, and running FramerAI without switching it on behaves exactly as it did before.

   Memory. Each experience is stored along with how important it seemed and how the system felt at the time. Unused memories fade by half over time; recalling one makes it stronger. When space runs out, the weakest memory goes, not the oldest — so something rare but important outlives something recent and dull. Looking something up weighs three things together: how similar it is, how recent, and how important. Recurring experiences get quietly compressed into general concepts, the way many individual trips to a café become the idea of a café.

   Curiosity. The system measures how surprising something is, and the part doing the measuring learns as it goes — so repetition becomes boring and genuinely new things stand out. Crucially, it chases what it is currently getting better at, rather than whatever is merely unpredictable. Random noise is unpredictable forever and teaches nothing; this distinction is what keeps curiosity from getting stuck staring at static.

   Mood. Five internal dials — pleasantness, alertness, confidence, curiosity, and tiredness — drift back toward a resting level and get pushed around by what happens. They are not cosmetic: they change how adventurous or conservative the model's answers are. An internal state that never reaches the output is decoration; this one has an effect you can see.

   Sleep. Past a tiredness threshold, the system replays its own memories, favouring the important ones. It rehearses them, forms new concepts, drops weak memories, and writes a short first-person reflection on what happened. Replayed experience can also be fed back into training, so the sleep cycle can genuinely teach it something.

   Live senses. It can watch a camera and listen to a microphone continuously, either in the foreground or quietly in the background. A filter keeps it from recording the same thing over and over: it only pays attention when the input actually changes, with a forced look every so often in case it is missing something. Without that filter, a camera running at two frames a second would bury the memory in thousands of identical entries. Camera and microphone support are optional — the same machinery runs with no hardware attached, and it reports honestly whether it genuinely looked at something or only skimmed it.

   Languages. It identifies the writing system first, then narrows down which language it is by looking at common short words. If it cannot tell, it says so with low confidence rather than guessing English. Every memory records what language it was in, and competence is tracked separately per language — so being fluent in one does not get mistaken for being fluent in another. Word counting works for languages that do not put spaces between words. When it is confident, it asks for a reply in the same language.

   Using it. Once switched on, conversations run through this layer, and every reply comes back with a record of what was recalled and how it was feeling. There are commands to show it something, play it a sound, watch a video, run a live session, prompt it to wonder or reflect, give it feedback, and ask it to describe its own state. Everything is saved after each exchange.

   No new software dependencies were required. 156 automated tests cover memory fading and eviction, boredom, curiosity, mood limits, sleep, the change filter, seeing and hearing, running safely across threads, language identification, saving and loading, and the live commands.

   To be explicit: this does not claim FramerAI is conscious, and the code says so where it matters. These are working analogues of memory, curiosity, mood, and sleep — each one visible in a trace and tested on its own terms.

3. Aug 8, 2023Milestone

   The two-trillion-parameter model↗

   FramerAI got a new flagship size: a single model with two trillion parameters that handles text, code, images, video, and audio together, rather than five separate specialist models bolted to each other.

   A parameter is one of the adjustable numbers the model learns during training — roughly, the raw capacity it has to store what it knows. Two trillion of them is the size of the whole thing on disk. But the model only wakes up a small slice for any given word it reads or writes: about 49 billion, or one fortieth of the total. This is what makes a model this large practical to run. The design is a bit like a large hospital: two thousand specialists on staff, but any one patient only sees the four they actually need.

   The two trillion breaks down as roughly 1.96 trillion for the language core and 32 billion for the parts that see, hear, and generate pictures, video, and sound. Storing it takes about 3.7 terabytes; training it needs about 29 terabytes of working space, since training keeps several extra copies of every number in flight.

   Two design decisions in this entry are worth spelling out:

   The model has 384 specialists rather than a smaller number of larger ones. Same total size either way, but more specialists means a smaller fraction is active at once, which is cheaper to run. The number 384 was chosen because it divides evenly across 8, 16, 32, 64, or 128 machines — so the specialists can be split across a datacenter without leftovers. That was settled before the code to actually split them existed, because it is far more painful to change later.

   The image resolution setting was deliberately left small. A later change would let the model read a large picture by cutting it into tiles, and this setting was about to become the size of one tile rather than the size of the whole picture. Raising it would have locked in a fixed resolution and defeated the tiling before it was written.

   Checking the size of the model costs nothing — it is arithmetic, not construction — so this still runs on a laptop and in automated testing.

   Alongside the model itself, the size figures were locked down with automated checks. Later work would swap out entire components, and without those checks the published numbers could quietly drift away from reality.