Real project status
The first Crisol already breathes.
This is not a paper. The first Crisol Mini has been trained from scratch: 5,000 steps, without a single NaN or Inf, the loss falling from 45 to a best of 3.5 and causal reasoning switching on by itself. All in 24 hours and for ten cents of electricity. Here is the real status — no makeup, including what is still missing.
A model trained from scratch, not a borrowed fine-tune.
The first Crisol Mini — ~2.13 B parameters, a single expert per layer — was forged entirely on its own architecture: 12 layers, 5 universal slots, an O(1)-cost HoloBinder, holo_dim 4096, a NAR of 2048 axes and a NOE of 2048 dimensions. Not a single weight inherited from an external model.
Across 5,000 steps it produced not a single NaN or Inf: bounded gradients, healthy distributions, stable training from start to finish. And as the loss fell from 45 to a best of 3.5, reasoning woke up on its own — causal confidence rose from 0 to ~0.8 and the geometry aligned with the NAR axes. It doesn't just predict: it begins to reason.
And its whole world already lives in its 64,000-piece vocabulary — the names of its agents, its validators and its memory structures:
Phased roadmap
From clean ground to an organism that breathes.
Six phases done, one in progress, one pending. The launch target is September 2026.
- Phase 0
Preparation
doneBackups, a dedicated working branch, selective cherry-picks and Done criteria per sub-step. Clean ground before raising anything.
- Phase 1
Five new components
doneCoreEncoder, MemoryExpert, IdentidadStore, CausalStore and SuenoCausalAutomatico. The pieces that turn a model into a persistent organism.
- Phase 2
Backend integration
doneAll five components stitched into the engine's stack, layers and enums. Not a script on top: part of the forward pass.
- Phase 3
REST endpoint + Crisol Studio
doneFull control from the interface. Every parameter configurable, every process traceable, real-time metrics over WebSocket.
- Phase 4
Pre-trainings
doneA 64,000-piece BPE tokenizer, a NAR of 2048 axiomatic axes and a NOE of 256 invariants. The substrate of reasoning and knowledge, before the corpus.
- Phase 5
Forging the base model
doneThe first Crisol Mini trained from scratch — ~2.13 B parameters, 5,000 steps without a single NaN. The loss fell from 45 to a best of 3.5 and causal reasoning switched on by itself. Forged after fixing the learning rate and all of Crisol Studio.
- Phase 6
The five universal packages
in progressGenerating the .crisolpkg experts that travel through the Custodia cloud and import like apps. The container-ship starting to load.
- Phase 7
E2E validation + launch
pendingEnd-to-end testing of the whole organism and public release. Target: September 2026.
Where Crisol does not yet compete.
Crisol is not a "GPT-killer". Saying otherwise would be dishonest. There are arenas where a frontier LLM, as of today, wins — and they deserve to be named plainly.
General fluency
A Crisol Mini does not write prose as polished or as broad as a frontier LLM trained on trillions of tokens. That is not its goal.
Encyclopedism
It does not try to know everything from memory. Knowledge arrives modularly, as .crisolpkg packages, not baked into the weights.
Multimodality
Today it reasons over and generates text. Vision, audio and other modalities are not part of v1.0.
Raw speed
A closed model served from a datacenter answers faster per request. Crisol trades milliseconds for sovereignty.
But it competes — and wins — where it truly matters.
Crisol is not playing to be one more, larger model. It is playing a different game.
The why behind the status
The status is honest because the thesis deserves it.
If you want to understand what motivates each phase of this roadmap — why a sovereign organism and not a larger model — the manifesto explains it in full.
Read the manifesto