All models › CohereLabs
tiny-aya-l2-thinker
CohereLabs · license to accept license · on Arena since 2 Sept 2026
Download New · waiting for Arena votes
Arena score—Not voted on Arena yet
Size3.4 Bparameters
4-bit download2.1 GBweights only, Q4_K_M
Max context—
Where to use it for free
- Download and run it at home: official weights · GGUF files for Ollama and LM Studio (you accept the license on Hugging Face first)
New: waiting for Arena votes
Released on 2 Sept 2026. Arena has not voted on this model yet. As soon as it does, the model moves into the ranking automatically, usually the next night. Until then you can still check its size, memory needs and speed below.
Speed on popular hardware
| Hardware | GPU memory | Fits? | Speed |
|---|---|---|---|
| RTX 3060 12 GB | 12 GB | Fits in GPU | 101 tok/s |
| RTX 4060 Ti 16 GB | 16 GB | Fits in GPU | 80 tok/s |
| RTX 3090 | 24 GB | Fits in GPU | 261 tok/s |
| RTX 4090 | 24 GB | Fits in GPU | 282 tok/s |
| RTX 5090 | 32 GB | Fits in GPU | 501 tok/s |
| RX 7900 XTX | 24 GB | Fits in GPU | 268 tok/s |
| 2 × RTX 3090 | 48 GB | Fits in GPU | 254 tok/s |
| 4 × RTX 3090 | 96 GB | Fits in GPU | 240 tok/s |
| Mac M4 Max 128 GB | 96 GB | Fits in GPU | 153 tok/s |
| Ryzen AI Max+ 395 128 GB | 96 GB | Fits in GPU | 72 tok/s |
| 8 × H100 SXM | 640 GB | Fits in GPU | 773 tok/s |
Estimate at 4-bit with 32K context, 64 GB of DDR5 for offload. Speed ≈ memory bandwidth × 70% ÷ data read per token; real numbers vary by about ±30%.
GPU memory needed
| Level | 8K context | 32K context | 128K context | Quality vs full |
|---|---|---|---|---|
| Q3_K_M | 2.4 GB | 2.7 GB | 4.1 GB | +10.5% noticeable loss |
| Q4_K_M | 2.8 GB | 3.1 GB | 4.4 GB | +2.8% small loss |
| Q5_K_M | 3.1 GB | 3.5 GB | 4.8 GB | +0.9% very small loss |
| Q6_K | 3.5 GB | 3.8 GB | 5.2 GB | +0.3% near-lossless |
| Q8_0 | 4.3 GB | 4.7 GB | 6.0 GB | ≈ 0% same as full |
| F16 | 7.5 GB | 7.8 GB | 9.2 GB | ≈ 0% same as full |
Weights + context cache (from the model's architecture) + runtime. Quality: increase in perplexity against the full model, from llama.cpp tests, scaled by model size. Simulate your own setup.