Model
Model explorer

Aya 101

OPEN
Cohere Labs · Aya family · released Feb 12, 2024

13B mT5-based instruction model covering 101 languages, built by 3,000+ researchers; Apache-2.0, no hosted inference provider.

ReasoningCodingVisionFunction callingTool useAgentic
-349.2
Elo · unrated
Parameters
13B
Active params
13B (dense)
Context
1K tokens
Architecture
Dense encoder-decoder Transformer (mT5-XXL backbone, instruction-finetuned)
License
Apache-2.0
Languages
101+
API price (in/out)
No hosted API
Modalities
text
Benchmark results
Bar shows position within the tracked field; marker = field best
MMLUKnowledge37.3%#261
best: OpenAI o3 · 92.9%
XWinogradReasoning70.7%#6
best: Sarashina2-70B · 91.8%
Run it locally
VRAM @ Q4
—
VRAM @ FP16
26 GB
Fits on (Q4)
Multi-node cluster required
Throughput data unavailable.
Quantizations
—
Fine-tune it
Permissive
QLoRA10.2 GB1× RTX 3060 12GB
LoRA29.0 GB1× RTX 5090 32GB
Full fine-tune210.0 GB2× H200 141GB
QLoRA SFT on ~10k samples ≈ $7.61 (1× RTX 3060 12GB)
Plan a fine-tune →
Aya family
Elo progression across releases
Aya 101Feb 2024-349.2
Resources
HFCohereLabs/aya-101DOCScohere.com/research/aya
API price weights · each benchmark row carries its own source badge (see methodology)