Verified · Nov 7, 2026
Independently verifiedUnderdog made a playable browser game for its 31.7 MB speech model
2 sourcesUnderdog AI's own X post (opened this run via the dual oembed/syndication channels, post text agreeing word-for-word with the trailing media link rendering differently; created_at 2026-11-05T04:44:59.000Z, snowflake-confirmed) states: "We made a game for our new frontier text-to-speech model, Woof. 31.7 MB, so small that it runs in your browser" and "Play the game on X... all voices are Woof. Defeat Darkcloud for an invite" — "frontier" is the company's own characterization of its own model. TechCrunch (Ivan Mehta, article:published_time 2026-11-06T13:02:16+00:00; opened and read in full this run) frames it: "One such startup, Underdog, has released a game built with its new voice model called Woof that’s only 32MB large." — the headline's 32MB is TC's rounding of the post's 31.7 MB. The outlet adds: "All of this works within the browser through Underdog’s speech generation model, which, the company says, can run on most devices without any internet connection. The model has 31 million parameters and is open for everyone to download." TC's own hedge travels with the capability claim: "While such a small model likely won’t replace bigger models, it may still be enough to drive local assistant experiences or create personalized podcasts." Context, per TC: "This demo comes weeks after Underdog launched Bark, a 167MB speech-to-text model also designed to run on edge devices for transcription and dictation." And the founder: "Sigil Wen, the founder of Underdog, told TechCrunch that he started building the app because, instead of using notetakers like Granola, he wanted to build his own system with AI models that can run on users’ devices and let them own their data."
Why now
The demo post went up November 5 and TechCrunch's writeup landed November 6 — the post's own call to action is "Play the game on X", and per TechCrunch's description, typing a sentence into the in-game typewriter makes Woof speak it back. For creators tracking the edge-AI wave, TC's own hedge sets the frame: "While such a small model likely won’t replace bigger models, it may still be enough to drive local assistant experiences or create personalized podcasts."
Why it is worth publishing
A 31.7 MB speech model that runs in the browser is a tangible edge-AI demo a creator can actually play on camera — the post's own call to action is to play it, and TechCrunch's description highlights the in-game typewriter and the music tuner — the card keeps every capability claim welded to its source: the size and browser claim to Underdog's own post, the parameter count and download availability to TechCrunch, with TC's hedge carried in the same sentence.
Evidence basis
Two opened sources: Underdog's own X post (November 5, 2026, created_at 04:44 UTC, snowflake-confirmed) stating the model name, the size ("31.7 MB"), and the browser-runnability claim; and TechCrunch (November 6; article:published_time 2026-11-06T13:02:16+00:00) adding the reported parameter count (31 million), download availability, the game description, the Bark context (167MB speech-to-text, weeks earlier), and the founder quotes. No benchmark numbers exist in either opened source — the only capability evidence this record holds is the playable game itself, which was not opened or played this run.
“Underdog's new speech model is 31.7 MB small — it runs in your browser, and the company made a game for it where every voice is Woof.”
Angle
The edge-AI counter-trend card: a 31.7 MB speech model that runs in the browser, wrapped in a game the company says to play — what's company-stated, what's TechCrunch-reported, and what a creator can verify by playing it.
Format
Short talking-head video
Demo idea
Play the game on screen from the link in Underdog's post: type a sentence on the in-game typewriter and let Woof speak it back, rotate the tuner for the music, and say the size claim on camera — 31.7 MB per Underdog's post. If you download the model, show the file size on your own machine and label it as your own observation; keep TechCrunch's 31-million-parameter and download-availability statements attributed to TechCrunch.
Platform notes
"frontier text-to-speech model" is Underdog's own characterization of its own model — say the company calls it frontier, never state it as an established benchmark fact. The 32MB headline figure is TechCrunch's rounding; the post itself says 31.7 MB — carry 31.7 MB in your own voice or attribute 32MB to TC. "can run on most devices without any internet connection" is company-says per TechCrunch; the 31-million-parameter count and "open for everyone to download" are TC's reporting. TC's own hedge is part of the fact: a small model "likely won’t replace bigger models". No benchmark numbers exist in either opened source — don't compare Woof to big-lab TTS models. And the game was not played this run — play it yourself before you demo, and label your playthrough as your own observation.
Usable claims
- On Thursday, November 5, 2026 (created_at 2026-11-05T04:44:59.000Z, snowflake-confirmed), Underdog AI announced on X: 'We made a game for our new frontier text-to-speech model, Woof. 31.7 MB, so small that it runs in your browser' and 'Play the game on X... all voices are Woof. Defeat Darkcloud for an invite' — 'frontier text-to-speech model' is the company's own characterization of its own model. TechCrunch (Ivan Mehta, article:published_time 2026-11-06T13:02:16+00:00) frames it: 'One such startup, Underdog, has released a game built with its new voice model called Woof that’s only 32MB large.' — the headline's 32MB is TC's rounding of the post's 31.7 MB. The outlet adds reported detail: 'All of this works within the browser through Underdog’s speech generation model, which, the company says, can run on most devices without any internet connection. The model has 31 million parameters and is open for everyone to download.' TC's own hedge is part of its sentence: 'While such a small model likely won’t replace bigger models, it may still be enough to drive local assistant experiences or create personalized podcasts.' Context, per TC: 'This demo comes weeks after Underdog launched Bark, a 167MB speech-to-text model also designed to run on edge devices for transcription and dictation.' The founder, per TC: 'Sigil Wen, the founder of Underdog, told TechCrunch that he started building the app because, instead of using notetakers like Granola, he wanted to build his own system with AI models that can run on users’ devices and let them own their data.' And Wen's quote: 'I think we will be able to run most capable models of today on the devices you already own within months. We just want to train the best open models that are American-made, that are open, that are capable and allow people to have private AI,' he said, referring to his manifesto.
Evidence pipeline
From the news
Breakdown
Two opened sources, tier-split. Underdog's X post is the company's own claim: the model name, "31.7 MB", "so small that it runs in your browser", and the game framing — "frontier" is the company's own word for its own model. TechCrunch adds what the post doesn't carry: 31 million parameters, download availability, the device/offline claim (as company-says), the game description, the Bark context, and the founder quotes — all TC-attributed. The record's honest center: no benchmark numbers exist in either opened source, the game itself was not played this run, and TC's own hedge ("likely won’t replace bigger models") is part of the fact — so the card's only demonstrated capability is the one a viewer can check by playing it.
Sources
Risks
- Play the game yourself before you demo it on camera — the t.co link in Underdog's post leads to it, and your own playthrough is the only check this record cannot do for you; label the experience as your own observation. Say the size as 31.7 MB per Underdog's post (or attribute 32MB to TechCrunch's headline), attribute the 31-million-parameter count and download availability to TechCrunch, and keep the offline capability as company-says. When you describe the model's quality, carry TC's hedge ('likely won't replace bigger models') in the same sentence, and never benchmark Woof against big-lab TTS models — no opened source provides numbers for that comparison.
Demo ideas
- Play the game live on camera from the link in Underdog's post: type a sentence on the typewriter, let Woof speak it back, and say the 31.7 MB size claim.
- A side-by-side demo: your regular cloud TTS versus Woof in the browser — no benchmark claims, just your own observed latency, labeled as your own test.
- An explainer card on the edge-AI trend: Woof at 31.7 MB (Underdog's post) next to Bark at 167MB speech-to-text (TechCrunch), each figure attributed.