Tiny Aya Base 32K vs Tiny Aya En-Thinker: Parent and Child, Not Rivals

Two Hugging Face repositories, four safetensors shards each, and the shard sizes match to the byte — 1,998,698,496 / 1,996,494,056 / 1,996,494,088 / 708,852,792. Tiny Aya Base 32K and Tiny Aya En-Thinker are not two laboratories' answers to the same question. They are the same 3.35B Cohere2 architecture at two different stages, and Cohere Labs' own model card says so out loud: the base checkpoint "is used as the base model for Tiny Aya L2-Thinker and Tiny Aya En-Thinker."

That makes the usual head-to-head framing wrong. You are not choosing between two competing 3B multilingual models. You are choosing between a starting point and a finished one — and the interesting question is what the post-training step between them actually bought, what it cost, and whether either is usable for what you have in mind given that both sit behind the same non-commercial licence and neither has a single published benchmark.

The one sentence that settles the relationship

Normally a "vs" piece has to infer the relationship between two checkpoints from architecture and parameter counts. Here you do not have to.

Tiny Aya Base 32K's card states it plainly, and it is worth reading carefully because of where the sentence sits — not in a footnote, but in the model summary, between the description of the checkpoint and the developer credits:

"This is a base pretrained model and has not been instruction-tuned or preference-aligned. This checkpoint is used as the base model for Tiny Aya L2-Thinker and Tiny Aya En-Thinker."

What makes that worth a paragraph is the inconsistency it exposes. En-Thinker's own card does not name Base 32K. Its closing line sends readers to "tiny-aya-global, tiny-aya-base, tiny-aya-l2-thinker" — and tiny-aya-base is the February checkpoint, documented at 8K context, not the 32K variant that says it is En-Thinker's parent. En-Thinker claims 32K on its own card. A model cannot be post-trained into a window four times wider than the window it was pretrained with. So the 8K base was never going to be the answer, and the 32K base's claim fits the arithmetic where En-Thinker's own pointer does not.

The likely explanation is mundane: Base 32K was uploaded on September 8, 2026, six days after the Thinkers, so En-Thinker's card was written before its parent existed and has not been updated since. That is inference from timestamps, not a vendor statement — but it is the reading that requires the fewest assumptions, and it means the newest document in the family is the most informative one.

Dimension by dimension

Everything below is what the two cards state, minus the duplication — both are 3.35B, BF16, text-only, Cohere2ForCausalLM, CC-BY-NC-4.0, gated, and unbenchmarked.

• Stage — Tiny Aya Base 32K is a pretrained base, "not instruction-tuned or preference-aligned"; Tiny Aya En-Thinker is post-trained on multilingual reasoning data with English reasoning traces.

• Chat template — Base 32K ships no chat_template.jinja at all; En-Thinker ships one, and its card devotes a whole section to how it renders turns.

• Prompting style — Base 32K's own example is completion (prompt = "The capital of Spain is"); En-Thinker expects apply_chat_template() with a messages list.

• Reasoning mode — Base 32K has no reasoning mode; En-Thinker is dual-mode, appending /think by default and emitting a thinking trace in English, or /no_think with enable_thinking=False for a direct answer.

• Language scope — Base 32K's card claims training on 70+ languages and names 67 explicitly; En-Thinker covers "44 languages plus English" with English reasoning traces, extending to 20+ more through additional non-reasoning instruction data.

• Architecture footprint — identical. Same four shard sizes, same parameter count, same config file length.

• Context — 32K input plus output on both cards. Neither is verifiable: both configs are behind the licence gate.

• Benchmarks — none on either card. No third-party evaluation of either model exists.

• Traction — Base 32K shows zero downloads and one like; En-Thinker shows seven downloads and two likes. Neither appears in an inference provider's catalogue.

What the post-training step bought

Three things, all of them behavioural rather than structural.

The first is a usable interface. A checkpoint with no chat template is not a model you prompt; it is a model you continue. Every conversation you have with Base 32K has to be serialised into text yourself, and the card says as much: "prompts should use text completion rather than a chat template. Additional post-training is recommended before deploying it as a conversational assistant." En-Thinker has already made that decision for you, down to the token layout.

The second is reasoning as a switch. En-Thinker's /think and /no_think modes let the same weights either produce a visible chain of thought between thinking tags or answer directly, and the card documents passing a prior thinking block back into the conversation as an assistant turn. That is a post-training artefact — there is nothing in a base checkpoint that responds to it.

The third is the design bet at the centre of En-Thinker: that reasoning happens in English even when the question and the answer are in another language. The card is explicit that this distinguishes it from Tiny Aya L2-Thinker, "which thinks in the same language as the prompt." Whether planning in a high-resource language and answering in a low-resource one actually helps is an open research question, and En-Thinker exists to let people test it rather than to settle it. Base 32K, having no reasoning mode at all, is silent on the question — which is precisely why it is the right control for anyone trying to answer it.

What the post-training step cost

The honest answer is that nobody can quantify it, because neither model has been evaluated on anything. But the two cards together hint at where the trade-off lives, and it is not where you would expect.

It is not language coverage. En-Thinker's narrow 44-plus-English reasoning scope is a property of its reasoning training data, and the card says coverage extends to 20+ further languages "through additional non-reasoning instruction data" — so the model still handles them, just without the thinking path. It is not the context window either; both claim 32K.

It is breadth of behaviour. Base 32K's card lists "continued pretraining, instruction tuning, and research on multilingual and long-context language modeling" as intended uses, with multilingual text generation, summarization, translation and cross-lingual adaptation all named as downstream applications. En-Thinker's card is narrower: "mathematics, science, and general reasoning tasks, as well as instruction following and multilingual open-ended generation." A post-trained checkpoint is opinionated in a way a base checkpoint is not, and the limitation Base 32K's card flags — "Chain-of-thought reasoning tasks such as multilingual math are comparatively weaker" — is exactly the weakness post-training set out to fix.

So the family reads as a division of labour rather than a competition. The base is broad and unfinished; En-Thinker is narrow and finished; and the base card's own text calls it what it is — the substrate both Thinkers were shaped from.

The licence is the constraint that actually binds

Both models are CC-BY-NC-4.0 with Cohere Labs' Acceptable Use Policy layered on top, both sit behind the same gate that asks you to accept terms and register before files download, and both route commercial questions to Cohere Sales.

This is worth stating before any technical comparison, because it decides the question for a large share of readers. If the plan is a commercial product, neither checkpoint is a candidate without a conversation with Cohere, and no amount of long-context behaviour or reasoning-mode flexibility changes that. Where non-commercial is genuinely fine — academic work, internal research, a benchmark harness, a comparison against your own model — the licence is irrelevant and the technical choice is the whole decision.

Neither has been benchmarked. Here is how to choose anyway

The absence of numbers is real and it is not going to be resolved by reading more carefully. Both cards make zero performance claims, no leaderboard lists either model, and neither has an inference provider behind it — meaning no provider has published the throughput or pricing that usually substitutes for a benchmark. What you can do is pick on structure, where the evidence is solid.

• Choose Tiny Aya Base 32K if you are going to train. Continued pretraining, instruction tuning, or a domain adaptation on top of a multilingual 3.35B model is what the card says it is for, and starting from a base checkpoint means you are not fighting post-training you did not choose. The 32K window also supports long-context research that the 8K February base could not.

• Choose Tiny Aya En-Thinker if you want to prompt something today. It is the only one of the two that will hold a conversation, and the only one with a reasoning mode at all.

• Choose both if the question is scientific rather than practical. The pair is a controlled comparison — same architecture, same size, same window, differing mostly in whether post-training happened — for asking whether English reasoning traces improve multilingual answers. That is the question En-Thinker was released to let people probe, and its own parent is the cleanest baseline.

One practical note on evaluating either: they are unproven research checkpoints with no deployment history, and a 6.7 GB BF16 download plus GPU time is a real cost to spend on a model that has never been independently run. Running that comparison through a single endpoint rather than standing up weights for each candidate is cheaper — OrcaRouter carries 195 models behind one API with 0% markup on provider list prices and automatic failover between providers, so a research checkpoint you are merely testing never sits on a production path. Tiny Aya is not on it and we do not host it; the point is only that the models you would test it against mostly are.

What would settle this

Two things, and both are cheap for Cohere Labs to publish and expensive for anyone else to produce.

The first is a benchmark — any benchmark. A single multilingual reasoning evaluation run on both checkpoints would convert the entire family from "card-stated" to "measured," and it would immediately answer whether the English-thinking design beats the same model without post-training, which is the research question the release is built around.

The second is a config. The 32K claim on both cards is unverifiable while the repositories are gated, and the difference between a genuinely long-context pretraining run and a positional-encoding extension applied to an 8K checkpoint is large in practice even though both produce a model that accepts 32K tokens. Opening the two config files would close that gap in a way no marketing copy can.

Until then the accurate answer to "Tiny Aya Base 32K or Tiny Aya En-Thinker" is that the comparison is real but the contest is not. Same weights, same window, same licence, same silence from everyone who has not downloaded them — one of them just has the chat template and the thinking tags, and the other is what it was made from.