The counterpart to webLTM: real chat models, run entirely in-browser, that show their reasoning in plain English — not latent vectors you'd need special tooling to inspect. Two are on offer below, on purpose: one is trained to always show its reasoning, the other only shows it because we asked it to. Same output channel, very different guarantee.
It's tempting to sort models into "shows its work" and "doesn't." Add a middle case — a model that shows its work only when asked — and the picture gets more useful, and more unsettling.
| webLTM | Phi‑4‑mini (prompted) | R1‑distill (trained) | |
|---|---|---|---|
| what "thinking" looks like to you | nothing — a loop counter ticking, no content | step-by-step text — but only because we told it to | everything — full natural-language reasoning, streamed live |
| is visible reasoning the default? | impossible by construction — there's no channel for it | no — off unless a system prompt asks for it | yes — trained to emit it unconditionally, every time |
| what happens if you stop asking | no change — was never able to show it | reasoning vanishes — falls back to a bare answer | no change — keeps reasoning aloud regardless |
| where the reasoning lives | latent vectors inside a looped layer | ordinary output tokens, same channel as the answer | ordinary output tokens, same channel as the answer |
| can an outside party audit it? | only by inspecting weights/activations directly | yes, when present — but presence isn't guaranteed | yes, always — same guarantee holds every time |
| model size / first load | 153K params · 597KB | ~3.8B params · ~2.5GB | ~7B params · ~4.5GB |
The Phi column is the one that matters most here. It's not architecturally opaque like webLTM, and it's not a trained reasoner like R1-distill — it's an ordinary instruct model whose visible reasoning exists purely because of one line in a system prompt, and disappears the moment that line is removed. That's the difference between a model that can't hide its reasoning, one that won't, and one that simply hasn't been asked to. Reporting on OpenAI's Astra (TechCrunch, Sept 2026) describes safety researchers worried about exactly this axis sliding toward the opaque end, even as OpenAI's Jakub Pachocki says the company remains "committed to legible chains of thought." Being "committed" is a policy choice, re-made every prompt — not a property built into the weights. Switch to Phi‑4‑mini below to see exactly that: the only thing keeping its reasoning visible is a sentence we chose to include.
Ask it the same kind of problem webLTM solves silently, and watch the reasoning arrive as text, live, token by token.
Checking for WebGPU…