The Mechanism

Thinking in neuralese.

RESEARCH 01

Latent Update Gate

Instead of always emitting a token, the model can pass its own hidden state forward as the next input through a trained update gate, with a separate head deciding when to stop and actually respond. That's the mechanism, and the loop is trained through, not just run at inference. Being precise about how much of it the released checkpoint actually got: for most of training a configuration bug meant only a single latent step ever reached the loss, so the depth was nominal. Of 82,697 training steps, 2,311 trained at a latent depth above one. That is the honest number, and it is small. What changed in those steps is still worth reporting — deeper thinking went from hurting to helping, and the best measured depth moved from two to four. The question that stayed open is the harder one: whether the latents carry reasoning specific to the question being asked. We re-ran that evaluation against the finished model. The trajectory is far more answer-directed than before, and the answer genuinely depends on it — but it is still not specific to the problem. Swapping in a different question's latent thoughts costs the model almost nothing.

Unscripted capture, not a mockup
STATUS

Progress so far

Latent reasoning at Cymela is early-stage research, but the checkpoint has moved. It now sustains multi-step latent computation: deeper thinking measurably helps, where in earlier runs it made things worse. The useful range is narrow — four steps is the measured optimum, and by eight the model does worse than if it hadn't thought at all — but inside that range the mechanism earns its cost, and it is cheap: at depth four the entire latent loop takes 0.69 seconds on a 2019 mid-range consumer GPU, with each additional step costing about 74 milliseconds. Thinking is not the expensive part. What it does not yet do is make those steps specific to the question. The model thinks, and the thinking is well formed, but you can hand it another problem's thoughts and it barely notices. Closing that gap is the next milestone, and it is the one that decides whether any of this was worth it.

Training Approach

Where the compute comes from.

Cloud TPU Access

Larger training runs use cloud TPU access (Kaggle TPU v5e-8). It's shared, free-tier compute — not a dedicated cluster — so runs are scheduled around availability.

Personal GPU Hardware

Day-to-day iteration — debugging, small evals, quick latent-reasoning tests — happens on personal AMD ROCm hardware, not a data center.

Debug, Rerun, Repeat

Runs fail, get debugged, and get rerun. That tight loop — not a big cluster — is what actually moves the latent-reasoning work forward.

Partnerships & Funding

Cymela is independently funded. If you're interested in supporting the research, a compute partnership, or another form of collaboration, reach out at contact@cymela.com.