←
Zellige AI14 posts
Zellige AI avatar

Zellige AI

@ZelligeAI

Intelligence, Tile by Tile. A small startup AI research company located in the US.

◷ Joined May 2026▧ 4 Photos and videos
Zellige AI avatar
Zellige AI@ZelligeAI·Nov 11

Tired of weekly limit reached? Monthly limit hit? Unable to finish your work day? Unclear usage limits? Zellige Kiln, You pay for your inference time. We have no weekly limits, No monthly either, no token counters, You get a certain number of hours each day of raw inference time, Every day that meter resets, History isn’t billed by the token, Time between requests is free, And in our simulated workloads solo plan, Covering your workday, Everyday, and more, all for $99/month. Too much? Use our build your own plan feature, Don't use all your hours? save money by lowering your allocated usage each day, A power user? Upgrade your concurrency or how many hours a day you get. Expect full privacy, We won't train on, store or sell your requests or the model's replies. Zellige Kiln is currently in waitlist, More details, features and waitlist signup at zellige.ai/pricing

0:36
Zellige AI avatar
Zellige AI@ZelligeAI·Nov 6

Tessera-27B-B-CJK-RL: our compressed-CJK coding agent, now trained with RL. AIME + HMMT 2026 11 point increase over non RL, Per agent turn it still uses 0.67x the base model's thinking. Open weights: huggingface.co/ZelligeAI/Tes… And our followup paper: zellige.ai/research/tessera-… Big thanks to @tinkerapi for sponsoring our continued research and development, We have much more in the works and have plans for more models and research in the near future.

Tessera-27B-B-CJK is a Qwen3.8-27B coding agent that thinks in compressed CJK, And saves tokens while doing so, Open weights, Apache-2.0. This is our first full tessera release, More info in the thread, check out our paper & hf: zellige.ai/research/tessera-… huggingface.co/ZelligeAI/Tes…

Zellige AI avatar
Zellige AI@ZelligeAI·Oct 14

This is first offical release of many models in the Tessera lineup, we have other ideas, research and optimizations currently in the works and plan to keep release not only models but other research and hopefully our own model line up at one point.

Zellige AI avatar
Zellige AI@ZelligeAI·Oct 14

Per agent turn, on 11 held-out tasks: - 0.50-0.64x base thinking tokens - 100% valid tool calls, 0 loops, tested at short and long context - Agentic and coding focused Our huggingface we also released ggufs, nvfp4 and the base LoRA adapter.

Zellige AI avatar
Zellige AI@ZelligeAI·Oct 14

Which we plan to explore and showcase more "B" being part of our beta release lineups as new and different variants testing and showcasing new novels ways of efficency can be applied to real models on real workloads rather then just on paper.

Zellige AI avatar
Zellige AI@ZelligeAI·Oct 14

This release and paper is part of our Tessera lineup, a set of models and tunes that will showcase novel and new ways of saving tokens, efficiency and other parts of our research, This model mainly focuses on the CJK thinking aspect and dissociation of thinking to output,

Zellige AI avatar
Zellige AI@ZelligeAI·Jul 15

A new technique to compress reasoning, Beyond caveman | Zellige AI We're Zellige, a small AI lab, and this is our first public release, a preview release at that, so we'll get into the flaws and our future plans below. This isn't our first paper but it is our first more public release, and we're trying to get more involved with the community over the coming months. Tessera-Preview-9B is OmniCoder-9B (Qwen3.5-9B hybrid) fine-tuned on ~10K examples where all the reasoning was machine-compressed into telegraphic CJK/symbol notation, while the visible output stays normal code and tool calls. Minimal example of the internal reasoning on a LiveCodeBench problem it solved: ``` 设 s 为 0-indexed 字符串。操作:选 i,c=s[i],删除 i 左侧最近 c 和右侧最近 c。 观察:操作不改变字符种类集合,仅减少计数。故最终长度 = 不同字符数。 算法:返回 len(set(s))。⊘:s 长度 1 → 返回 1。 ``` And the following response: ``` class Solution: def minimizedStringLength(self, s: str) -> int: return len(set(s)) ``` Now let's discuss benchmarks: LiveCodeBench-1055 pass@1 greedy: 34.9% vs 39.5% base IFEval: 40.9% vs 59.9% These are a regression, and we will get into why and its mostly not the thinking but importantly: Median output per problem: 639 tokens vs 16,384 The failure modes barely overlap. The base is less likely to write wrong code, it just drowns in its own thinking. Tessera finishes fast and cheap, and when it fails, it fails by writing wrong code. Part of the reason the dropoff occurs is that the training corpus is 100% code, zero instruction-following prompts (hence the preview label). We're releasing this and the paper now because we have much larger scaling plans: a 50-100x larger dataset and a model 4-5x larger for our next preview release. Now its time for the bread and butter of this model, How it actually works, and what CJK even is here: CJK just means the Chinese/Japanese/Korean character space, and in practice the register is mostly Chinese with math and logic symbols mixed in. The pipeline: take ordinary English reasoning traces, have a small compressor model rewrite them passage by passage into the telegraphic notation (with deterministic checks that every number and identifier survives, and a rules-only fallback when a rewrite fails a check), then fine-tune the 9B on the compressed thinks. Nothing is compressing at inference time. The model just learned to think that way natively. Why we think this works as well as it does: The compressed register is still understandable language, The base model already reads and writes Chinese from pretraining, so we're not asking it to invent a private encoding from 10K examples, we're asking it to think in a denser language it already knows. One Chinese character carries roughly a word of meaning, code identifiers and numbers stay verbatim, and what gets crushed is the connective filler ("okay, so now let me check whether..."). It also stays inspectable, unlike latent/continuous-reasoning approaches: you can read the thinks with any translator. Force-close the think block on the same weights and accuracy collapses from 66% to 4%, with 48 of 50 outputs coming back empty. We believe it to be a even better solution to the caveman idea that many other labs have been looking at. One note we want to disclose: CJK notation itself only buys ~1.7x (CJK chars cost ~1 token each, English amortizes ~4 chars per token, so the density win mostly evaporates at the tokenizer). The rest is that the training data teaches short thinks. A disciplined-English control is the obvious next experiment and it's in the paper's future work. We also released the 1.5B compressor that mints the training data, so other datasets can be converted into this notation and format of thinking. Future experiments with this system will also involve compressing chat history, since we've found other models can understand the compressed reasoning when asked to read it. We'll be around in the comments. If you poke at the weights and find something interesting (or broken), we want to hear it, This isn't our full drop yet, calling this even a preview is generous we have much more interesting things soon and scaling 5-50x this in the coming months.

Post attachment