Tiny Aya

Cohere Labs / Cohere · 2026-03-12 · analysed 2026-08-10 · skill v0.3.5

dense hybrid text-only open-weights medium on-device

Design overview

Tiny Aya is a 3.35B dense decoder-only Transformer (36 layers, hidden 2048, 16 Q / 4 KV heads, SwiGLU, bias-free, parallel attention+FFN blocks) whose block is deliberately carried over from Cohere's Command A: a 3:1 interleave of sliding-window attention (4096-token window, RoPE) and full attention (NoPE). The model's real design center of gravity is the data system around that block. A single 262k-vocabulary tokenizer is shared by all five releases, trained on 50GB of Fineweb-2 with a weighting that multiplies data-distribution weight by language-family/script bucket weight so that underrepresented scripts (Khmer, Telugu, Gujarati, Ge'ez) get competitive compression. Pretraining runs 6T tokens over 70 languages plus code, with a WSD schedule and a SmolLM3-style cooldown mixture. Posttraining is region-aware: five regional SFT clusters (Europe, West Asia, South Asia, Asia-Pacific, Africa) plus a global SFT model, fed by translated and FusioNN-aggregated synthetic data from Gemma3-27B-It, Command A and DeepSeek-V3 teachers, then each region-specialized checkpoint is merged with the global model under SimMerge operator selection — producing Tiny Aya Base, Global, Earth, Fire and Water from the same weights. The report's evaluation apparatus is as much a contribution as the model: rubric-based absolute judge ratings instead of win rates, language-confusion tracking, and mean+minimum safety reporting across languages. At 8K context it targets balanced multilingual quality rather than peaks, and its edge-deployment story (q4_k_m at 2.14 GB, ~10–32 tok/s on iPhones) is the closest the atlas has to an explicitly on-device massively-multilingual model.

What the report claims is novel

Atlas assessment: Architecturally a faithful, smaller Command A: parallel blocks, 3:1 sliding-window/full-attention interleave with the same RoPE-on-SWA / NoPE-on-full split, GQA, SwiGLU, bias-free — the report states this lineage explicitly, so the block itself contributes no new mechanism to the atlas. The novelty is entirely in the data and training system: the language-family/script-bucket tokenizer weighting (from Cohere's own prior work), region-clustered SFT with SimMerge-guided merging into Earth/Fire/Water variants, FusioNN teacher aggregation, and the rubric/language-confusion evaluation methodology. vs google-gemma-3 (the closest atlas peer at similar scale): Tiny Aya trades a uniform attention block for the SWA/NoPE hybrid, and its claim to distinctiveness is balanced cross-language performance and safety rather than raw peaks — consistent with its own framing that it does not 'nominally score the highest' on discriminative tasks. The family's on-device framing (quantized MLX/llama.cpp deployment) is also more explicit than most atlas peers at this scale.

Relation to the atlas

Lineage in

Influence out

No descendants recorded yet.

Notable omissions

Closest relatives in the atlas

Spec sheet

Identity

FamilyTiny Aya[1]
OrganisationCohere Labs / Cohere[2]
Release date2026-03-12[3]
Report URLhttps://arxiv.org/abs/2603.11510[4]
Licensenot disclosed
Open weightsyes[5]

Variants

Scale

Total parameters3.35B[6]
Active parameters / token3.35B
Layers36[7]
Hidden dimension2048[8]
FFN inner dimension11008[9]
FFN ratio5.375
Query heads16[10]
KV heads4[11]
Head dimension128
Vocabulary size262144[12]
Embedding tyingnot disclosed

Core block

Block typedense
Attention varianthybrid[13]
Attention layer pattern3:1 interleave of sliding-window attention (4096-token window) and full attention layers across the 36 layers, following Command A; sliding-window layers use RoPE, full-attention layers use NoPE[14]
Depth mixingsequential-residual
ActivationSwiGLU[15]
Expert countn/a[16]
Experts per tokenn/a[16]
Shared expertsn/a[16]
Routingn/a[16]
Load balancingn/a[16]
Expert granularityn/a[16]
Methodother[17]
RoPE base frequencynot disclosed
Partial RoPEnot disclosed
Typenot disclosed
Placementnot disclosed
QK-normno
Attention sinksno
Softcappingno
Othernot disclosed

Context

Trained context length8,192[18]
Deployed context length8,192[19]
Extension methodnot disclosed

Tokenizer

Algorithmnot disclosed
Notessingle massively multilingual tokenizer shared by all models; 262k vocabulary trained on 50GB of Fineweb-2 sampled with a weighting that combines data-distribution weight and language-family/script bucket weight (wi = wid·wib / Σ); GPT-4o regex pre-tokenization, no normalization; competitive or superior tokens-per-character compression across scripts, especially Khmer, Telugu, Gujarati, Lao, Ge'ez[20]

Training

Training tokens6T[21]
Data compositionlarge corpus of public and proprietary sources covering 70 languages plus programming-language datasets; filtering pipeline: (1) language ID and stopword filtering, (2) heuristic cleaning from raw sources, (3) deduplication, (4) domain classification and quality filtering; cooldown (mid-training) mixture upsampling the highest-quality pretraining datasets plus instruction-style datasets spanning all 70 languages[22]
Curriculum / stagingWarmup-Stable-Decay (WSD) schedule with a cooldown (mid-training) mixture following SmolLM3-3B; hyperparameters chosen via 200B-token ablation runs with 40B-token cooldowns[23]
Optimizernot disclosed
LR schedulepretraining: Warmup-Stable-Decay (WSD); posttraining (all cluster models): cosine decay, peak 2.5e-5, final 1.2e-6[24]
Batch schedulenot disclosed
PrecisionFP8 training mixing FP8, BF16 and FP32: main weights and optimizer states in FP32, weights cast to BF16/FP8 before computation; exponentials, softmaxes, layer norms and output embeddings in FP32; attention computation in BF16[14]
Parallelismnot disclosed
Hardware256 NVIDIA H100 GPUs for pretraining; 16 NVIDIA H100 GPUs per cluster model for posttraining (each cluster run completes within 24 hours wall-clock)[24]
Disclosed computenot disclosed

Post-training

SFTyes[25]
Preference optimisationother[26]
Reasoning trainingnone disclosed
Distillationnone

Modality

Typetext-only
Attachmentn/a[27]

Inference efficiency

KV-cache designGQA with 16 query heads over 4 KV heads: KV-cache footprint proportional to 4 heads (1/4 of MHA); sliding-window layers need only the 4096-token window cached
Quantisation shippedllama.cpp formats q4_0, q4_k_m and q8_0, served via llama.cpp and MLX; Q4_K_M optimal: 2.14 GB memory, 32.4 tokens/s decode, ~1.4-point mDolly degradation (Q4_0: ~2.1 points; Q8_0: negligible)[28]
Speculative decodingnone disclosed[29]
Serving optimisationsllama.cpp and MLX inference stacks for edge deployment; ~10 tokens/s decode on iPhone 13 and ~32 tokens/s on iPhone 17 Pro (Q4_K_M); low prefill throughput on older devices raises TTFT[28]

Evaluation

Benchmarks (report's own numbers only)

BenchmarkValueRef
Global MMLU44.9Table 4
INCLUDE45.1Table 4
Global PIQA68.3Table 4
mDolly (66 languages)86.9Table 5
mArenaHard (66 languages)67.4Table 5
GlobalMGSM (35 languages)52.8Table 5
Flores (66 languages, ChrF)43.5Table 5
WMT24++ (55 languages, ChrF)46.0Table 5
MultiJail Mean Safe Rate91.1Table 7
MultiJail Min Safe Rate87.0Table 7
XSTest Over-Refusal10.4Table 7
XSTest Under-Refusal15.5Table 7

Provenance references

  1. stated · Title
  2. stated · Author affiliations (Cohere Labs; Cohere)
  3. stated · arXiv stamp 'arXiv:2603.11510v1 [cs.CL] 12 Mar 2026'
  4. stated · pipeline
  5. stated · §1 ('a family of efficient, open-weight multilingual models'), §7 ('We release a family of 3.35B-parameter models: Tiny Aya Base, Tiny Aya Global, and region specific models Tiny Aya Water, Tiny Aya Earth, Tiny Aya Fire')
  6. stated · Table 2 (Total parameters 3.35B)
  7. stated · Table 2 (Num layers 36)
  8. stated · Table 2 (Embedding dims 2048)
  9. stated · Table 2 (FFN hidden dims 11008)
  10. stated · Table 2 (Num heads 16)
  11. stated · Table 2 (Num KV heads 4), §3.1 (grouped-query attention)
  12. stated · §2.1 ('a vocabulary size of 262k'), Table 2 (Vocab size 262k)
  13. stated · §3.1 (interleaved sliding-window and full attention layers in a 3:1 ratio), Table 2 (Sliding window 4096)
  14. stated · §3.1
  15. stated · §3.1 (SwiGLU activations, no bias in dense layers)
  16. n/a · §3.1 (dense decoder-only Transformer)
  17. stated · §3.1 (SWA layers use RoPE; full-attention layers use NoPE)
  18. stated · Table 2 (Input Context (tokens) 8192)
  19. stated · Table 2 (Input Context (tokens) 8192, Output Context (tokens) 8192)
  20. stated · §2.1
  21. stated · §3.1 ('We pretrain Tiny Aya model for 6T tokens')
  22. stated · §2.2
  23. stated · §3.1, §2.2
  24. stated · §3.1, §3.2
  25. stated · §2.3, §3.2 (region-specific SFT on five regional data subsets plus a global SFT model over all regions; 3 epochs, cosine decay, global batch 32)
  26. stated · §3.2 ('a minimal preference tuning phase on top of SFT for the Tiny Aya Global model... teaches the model its identity... while maintaining multilingual safety')
  27. n/a · §1, §4 (text-only model)
  28. stated · §6
  29. stated · §6 (no speculative decoding mentioned)