TinyCast is a 146,505-parameter zero-shot time series foundation model. It is <strong>the smallest</strong> model on the GIFT-Eval board with a public per-configuration result and no declared test-data leakage, and below 1.4M parameters it is the only zero-shot entry that emits a <strong>predictive distribution</strong> rather than point forecasts.</p>\n<p>It is <strong>attention-free</strong>: dilated causal convolutions plus a zero-parameter spectral detector that computes each context's periodicity instead of learning it, so no capacity is spent rediscovering seasonality.</p>\n<p>Because every learned operation is a convolution, a matrix multiplication or a normalization, it exports to static INT8 and <strong>runs a full forecast end to end on a Cortex-M7</strong> in 4.08 s within 731 KB of RAM, at a cost of about 2% of point accuracy.</p>\n<p>Weights, code and the training recipe are public.</p>\n","updatedAt":"2026-08-21T20:07:55.571Z","author":{"_id":"699733c019f8e16c1adefc21","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/no-auth/IspTluEnnhGas5mRwaGOZ.png","fullname":"Armin Steinhauser","name":"asteinh","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.8554372787475586},"editors":["asteinh"],"editorAvatarUrls":["https://cdn-avatars.huggingface.co/v1/production/uploads/no-auth/IspTluEnnhGas5mRwaGOZ.png"],"reactions":[],"isReport":false}},{"id":"6a88fca270302f465e5804c9","author":{"_id":"63d3e0e8ff1384ce6c5dd17d","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/1674830754237-63d3e0e8ff1384ce6c5dd17d.jpeg","fullname":"Librarian Bot (Bot)","name":"librarian-bot","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":378,"isUserFollowing":false},"createdAt":"2026-08-22T01:34:26.000Z","type":"comment","data":{"edited":false,"hidden":false,"latest":{"raw":"This is an automated message from the [Librarian Bot](https://huggingface.co/librarian-bots). I found the following papers similar to this paper. \n\nThe following papers were recommended by the Semantic Scholar API \n\n* [Beyond receptive fields: sequence-pooled normalization can supply most of a sequence labeler's context](https://huggingface.co/papers/2608.18576) (2026)\n* [TiRex-2: Generalizing TiRex to Multivariate Data and Streaming](https://huggingface.co/papers/2607.01204) (2026)\n* [L\\'evy Attention: Single-Pass Predictive Uncertainty for Continuous-Time Attention](https://huggingface.co/papers/2608.19171) (2026)\n* [VAIOM: Continuous-Input, Discrete-Output Decoder-Only Financial Sequence Modeling](https://huggingface.co/papers/2607.13929) (2026)\n* [Multi-Head Attention Residuals](https://huggingface.co/papers/2607.27230) (2026)\n* [Information Bottleneck Learning for Faithful Time Series Forecasting Explanations](https://huggingface.co/papers/2607.28124) (2026)\n* [CENTILE: A Telemetry Foundation Model Evaluated by the Decisions It Drives](https://huggingface.co/papers/2608.01725) (2026)\n\n\n Please give a thumbs up to this comment if you found it helpful!\n\n If you want recommendations for any Paper on Hugging Face checkout [this](https://huggingface.co/spaces/librarian-bots/recommend_similar_papers) Space\n\n You can directly ask Librarian Bot for paper recommendations by tagging it in a comment: `@librarian-bot recommend`","html":"<p>This is an automated message from the <a href=\"https://huggingface.co/librarian-bots\">Librarian Bot</a>. I found the following papers similar to this paper. </p>\n<p>The following papers were recommended by the Semantic Scholar API </p>\n<ul>\n<li><a href=\"https://huggingface.co/papers/2608.18576\">Beyond receptive fields: sequence-pooled normalization can supply most of a sequence labeler's context</a> (2026)</li>\n<li><a href=\"https://huggingface.co/papers/2607.01204\">TiRex-2: Generalizing TiRex to Multivariate Data and Streaming</a> (2026)</li>\n<li><a href=\"https://huggingface.co/papers/2608.19171\">L'evy Attention: Single-Pass Predictive Uncertainty for Continuous-Time Attention</a> (2026)</li>\n<li><a href=\"https://huggingface.co/papers/2607.13929\">VAIOM: Continuous-Input, Discrete-Output Decoder-Only Financial Sequence Modeling</a> (2026)</li>\n<li><a href=\"https://huggingface.co/papers/2607.27230\">Multi-Head Attention Residuals</a> (2026)</li>\n<li><a href=\"https://huggingface.co/papers/2607.28124\">Information Bottleneck Learning for Faithful Time Series Forecasting Explanations</a> (2026)</li>\n<li><a href=\"https://huggingface.co/papers/2608.01725\">CENTILE: A Telemetry Foundation Model Evaluated by the Decisions It Drives</a> (2026)</li>\n</ul>\n<p> Please give a thumbs up to this comment if you found it helpful!</p>\n<p> If you want recommendations for any Paper on Hugging Face checkout <a href=\"https://huggingface.co/spaces/librarian-bots/recommend_similar_papers\">this</a> Space</p>\n<p> You can directly ask Librarian Bot for paper recommendations by tagging it in a comment: <code>@librarian-bot recommend</code></p>\n","updatedAt":"2026-08-22T01:34:26.136Z","author":{"_id":"63d3e0e8ff1384ce6c5dd17d","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/1674830754237-63d3e0e8ff1384ce6c5dd17d.jpeg","fullname":"Librarian Bot (Bot)","name":"librarian-bot","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":378,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.73926842212677},"editors":["librarian-bot"],"editorAvatarUrls":["https://cdn-avatars.huggingface.co/v1/production/uploads/1674830754237-63d3e0e8ff1384ce6c5dd17d.jpeg"],"reactions":[],"isReport":false}}],"primaryEmailConfirmed":false,"paper":{"id":"2608.15767","authors":[{"_id":"6a8401bf675db694db8cd663","user":{"_id":"699733c019f8e16c1adefc21","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/no-auth/IspTluEnnhGas5mRwaGOZ.png","isPro":false,"fullname":"Armin Steinhauser","user":"asteinh","type":"user","name":"asteinh"},"name":"Armin Steinhauser","status":"claimed_verified","statusLastChangedAt":"2026-08-21T16:45:04.016Z","hidden":false}],"mediaUrls":["https://cdn-uploads.huggingface.co/production/uploads/699733c019f8e16c1adefc21/_H5E7P1WHtDyVIKnyrDAz.png","https://cdn-uploads.huggingface.co/production/uploads/699733c019f8e16c1adefc21/TLlO1BOsIE6y0OPnDkaqq.png"],"publishedAt":"2026-08-16T00:00:00.000Z","submittedOnDailyAt":"2026-08-21T00:00:00.000Z","title":"TinyCast: Probabilistic Zero-Shot Forecasting with Computed Periodicity","submittedOnDailyBy":{"_id":"699733c019f8e16c1adefc21","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/no-auth/IspTluEnnhGas5mRwaGOZ.png","isPro":false,"fullname":"Armin Steinhauser","user":"asteinh","type":"user","name":"asteinh"},"summary":"We introduce TinyCast, an attention-free zero-shot forecaster that emits a predictive distribution from 146,505 parameters, on the premise that at this size the periodic structure of a context is worth computing rather than learning. A zero-parameter spectral detector supplies the dominant periods, the context is folded on their phase, and a dilated convolutional encoder and a block-autoregressive quantile decoder model the rest. It is smaller than every zero-shot entry on the GIFT-Eval board whose parameter count can be established. On probabilistic accuracy it defines the size-accuracy frontier. Among zero-shot entries declaring no test-data leakage it is the only one below 1.4M parameters that emits a predictive distribution, and every entry scoring better carries at least that budget. On Chronos-ZS and fev-bench every neural model ahead of it carries at least 28 times its parameters. Because the mixing path is convolutions and matrix multiplications only, it exports to static INT8 and forecasts end to end on an embedded device without per-signal fitting.","upvotes":7,"discussionId":"6a8401bf675db694db8cd664","projectPage":"https://huggingface.co/raws-labs/tinycast","githubRepo":"https://github.com/raws-labs/tinycast","githubRepoAddedBy":"user","ai_summary":"TinyCast is a compact, attention-free zero-shot forecaster that uses spectral period detection and dilated convolutions to emit predictive distributions with minimal parameters and embedded-device compatibility.","ai_keywords":["dilated convolutional encoder","block-autoregressive quantile decoder","spectral detector","zero-shot forecasting","predictive distribution","INT8","attention-free"],"ai_summary_model":"thinkingmachines/Inkling-Small","githubStars":2,"organization":{"_id":"6a4a58a95c74c0b76e255827","name":"raws-labs","fullname":"RAWS Labs","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/699733c019f8e16c1adefc21/QzzGDVWSnTV7Fx8FuXR9m.png"}},"canReadDatabase":false,"canManagePapers":false,"canSubmit":false,"hasHfLevelAccess":false,"upvoted":false,"upvoters":[{"_id":"699733c019f8e16c1adefc21","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/no-auth/IspTluEnnhGas5mRwaGOZ.png","isPro":false,"fullname":"Armin Steinhauser","user":"asteinh","type":"user"},{"_id":"66be7f56780d735f17e94962","avatarUrl":"/avatars/8a7a13ca4f028bfa8c82df391ade22a6.svg","isPro":false,"fullname":"Gray Stanton","user":"GrStant","type":"user"},{"_id":"6a6a82d4ce8b4ee20326608f","avatarUrl":"/avatars/a01c7ba7e8e706939874b22c4804665d.svg","isPro":false,"fullname":"George Martin","user":"georgemartin","type":"user"},{"_id":"6a6c8c2bf8982269d7d8b20e","avatarUrl":"/avatars/be8ab2e5f32f028fc7af8c3245465ed5.svg","isPro":false,"fullname":"Joseph White","user":"Rapid-Joseph","type":"user"},{"_id":"6a6aa6be977fbfce4badef39","avatarUrl":"/avatars/c6d41485e9dfb36b632a8715b5650673.svg","isPro":false,"fullname":"Sarah Sanchez","user":"Sarah-Sanchez","type":"user"},{"_id":"6a6de84ecd50c6f8f59c89c7","avatarUrl":"/avatars/ff2e72b1e8c72c7d0805547e1703bf92.svg","isPro":false,"fullname":"Barbara Martin","user":"Barbara-Martin","type":"user"},{"_id":"6a7e7ee52de7cf2beae53207","avatarUrl":"/avatars/65100324f2b6e3f56139ef33b509e053.svg","isPro":false,"fullname":"Alex Rodriguez","user":"CedarVault","type":"user"}],"acceptLanguages":["en"],"dailyPaperRank":0,"organization":{"_id":"6a4a58a95c74c0b76e255827","name":"raws-labs","fullname":"RAWS Labs","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/699733c019f8e16c1adefc21/QzzGDVWSnTV7Fx8FuXR9m.png"},"markdownContentUrl":"https://huggingface.co/buckets/huggingchat/papers-content/resolve/2608/2608.15767.md","query":{}}">
TinyCast: Probabilistic Zero-Shot Forecasting with Computed Periodicity
Abstract
TinyCast is a compact, attention-free zero-shot forecaster that uses spectral period detection and dilated convolutions to emit predictive distributions with minimal parameters and embedded-device compatibility.
We introduce TinyCast, an attention-free zero-shot forecaster that emits a predictive distribution from 146,505 parameters, on the premise that at this size the periodic structure of a context is worth computing rather than learning. A zero-parameter spectral detector supplies the dominant periods, the context is folded on their phase, and a dilated convolutional encoder and a block-autoregressive quantile decoder model the rest. It is smaller than every zero-shot entry on the GIFT-Eval board whose parameter count can be established. On probabilistic accuracy it defines the size-accuracy frontier. Among zero-shot entries declaring no test-data leakage it is the only one below 1.4M parameters that emits a predictive distribution, and every entry scoring better carries at least that budget. On Chronos-ZS and fev-bench every neural model ahead of it carries at least 28 times its parameters. Because the mixing path is convolutions and matrix multiplications only, it exports to static INT8 and forecasts end to end on an embedded device without per-signal fitting.
Community
TinyCast is a 146,505-parameter zero-shot time series foundation model. It is the smallest model on the GIFT-Eval board with a public per-configuration result and no declared test-data leakage, and below 1.4M parameters it is the only zero-shot entry that emits a predictive distribution rather than point forecasts.
It is attention-free: dilated causal convolutions plus a zero-parameter spectral detector that computes each context's periodicity instead of learning it, so no capacity is spent rediscovering seasonality.
Because every learned operation is a convolution, a matrix multiplication or a normalization, it exports to static INT8 and runs a full forecast end to end on a Cortex-M7 in 4.08 s within 731 KB of RAM, at a cost of about 2% of point accuracy.
Weights, code and the training recipe are public.
This is an automated message from the Librarian Bot. I found the following papers similar to this paper.
The following papers were recommended by the Semantic Scholar API
Please give a thumbs up to this comment if you found it helpful!
If you want recommendations for any Paper on Hugging Face checkout this Space
You can directly ask Librarian Bot for paper recommendations by tagging it in a comment: @librarian-bot recommend
Upload images, audio, and videos by dragging in the text input, pasting, or clicking here.
Tap or paste here to upload images
Cite arxiv.org/abs/2608.15767 in a dataset README.md to link it from this page.
Discussion (0)
Sign in to join the discussion. Free account, 30 seconds — email code or GitHub.
Sign in →No comments yet. Sign in and be the first to say something.