Cohere Releases Tiny Aya L2-Thinker for In-Language Reasoning
The 3.35B model is designed to keep generated reasoning traces in the prompt language across a broad evaluation, while harder mathematics and independent validation remain open tests.
Edited by Tyronne Panaino
Cohere Labs published Tiny Aya L2-Thinker on October 6 alongside multilingual reasoning data for researchers studying how models generate intermediate text outside English. The accompanying preprint describes a 3.35-billion-parameter model and reports an in-language reasoning rate above 93% across 60 languages and six benchmarks. The release gives multilingual-AI researchers weights and data to inspect, but its performance evidence comes from the same team and has not been independently replicated in the sources reviewed here.
The central change is narrower than general multilingual support. Many models can answer in the language of a prompt while producing their visible reasoning trace in English. Tiny Aya L2-Thinker is trained to keep that generated trace in the prompt language as well. That may make the displayed process easier for more users to examine, although a readable trace is not proof that it faithfully represents a model's internal computation or that its final answer is correct.
The model separates task solving from reasoning language
The Cohere Labs research article describes a comparison between Tiny Aya L2-Thinker and an English-reasoning counterpart built from the same base. The final evaluation spans mathematics, commonsense and cultural reasoning, instruction following, and open-ended generation. The authors use the term L2 reasoning for a trace that stays in the language of the user's prompt.
That distinction matters because answer language and trace language are separate measurements. A system can return a fluent localized answer while routing its intermediate text through English. The release is therefore best read as a study of controllable trace language and cross-language transfer, not as evidence that every supported language receives equal factual accuracy, cultural coverage or safety.
Three kinds of data carry different parts of the job
The training approach combines English reasoning examples, a smaller multilingual reasoning set and multilingual examples that contain answers without reasoning traces. In the authors' account, English reasoning provides much of the task-solving backbone, translated multilingual reasoning teaches the desired trace language, and broader non-reasoning data helps that behavior transfer to languages without direct reasoning supervision.
The blog reports that adding the multilingual reasoning component moved the measured in-language rate from 12.8% to 86.1% in an intermediate comparison. Adding multilingual non-reasoning data then supported further generalization. The useful result is not that translation alone solves multilingual reasoning; it is that the composition of ordinary and reasoning data can change which language appears in a model's visible intermediate output.
What the reported 93% measures
The 54-page preprint defines its language-rate metric with automated language identification, using FastText and a GlotLID fallback. That method tests whether most of a trace is written in the prompt language. It does not, by itself, judge whether each step is logically valid, culturally appropriate or causally responsible for the final answer.
Across five of the six benchmarks, the Cohere article says task performance fell by no more than two to three points relative to the English-reasoning comparison, while the share of traces in the prompt language rose above 93%. Competition-level mathematics was the larger exception. This boundary is important: the release supports a broad language-control result, not a claim that all reasoning quality is preserved on every task.
Who should care and what to test next
For open-model researchers, the weights and data create a testable starting point for studying trace language without requiring a separate reasoning corpus for every target language. For teams evaluating multilingual assistants, the work suggests adding trace language, repetition and task accuracy to the same test plan rather than judging only the final response.
The next strong evidence would be independent reproduction across additional model families, native-speaker review of low-resource languages and evaluations that separate readable traces from faithful reasoning. Tests should also examine translation artifacts, culturally grounded tasks and the mathematics weakness described by the authors. Until those checks arrive, the release is a useful research artifact with bounded first-party evidence.
Status
Learning. The official article and author preprint document the model, released data, training design and reported evaluation. Internal confidence is medium because both sources come from the same research team, the paper is a preprint and no independent replication was fetched this run.
Sources
- Cohere Labs — Multilingual Bridges: How Data Mixing Unlocks In-Language Reasoning
- Mofakhami and co-authors — Building Multilingual Bridges preprint
Update note: Last reviewed 2026-10-13. We will revise this post if independent teams reproduce the results or the authors publish material evaluation changes.
Sources
Drafted with AI assistance from source briefs; reviewed for citation completeness and label accuracy.