Tri-21B
Tri-21B is an open-weights language model from Trillion Labs, released August 1, 2025. It has 21B parameters (dense).
We introduce Tri-21B, our flagship large language model that redefines the efficiency frontier in LLM training. By achieving state-of-the-art performance with only 2.3T training tokens, we demonstrate that exceptional capabilities don't require excessive computational resources.
Check it against your own PC →
Can you run Tri-21B locally?
The smallest common GPU it runs on is 8 GB of VRAM (with 32 GB of RAM): IQ3_XXS, split GPU / CPU, roughly 6 tokens per second. Estimates assume 32 GB of DDR5 RAM, a 16K context and llama.cpp-style GGUF quantization. Enter your exact hardware in the live app.
| GPU memory | Verdict | Best quant | VRAM | RAM | tok/s |
|---|---|---|---|---|---|
| 6 GB VRAM | Very slow | IQ3_XXS (noticeable loss) | 5.2 GB | 6.2 GB | ~4 |
| 8 GB VRAM | Runs, slowly | IQ3_XXS (noticeable loss) | 7.2 GB | 4.2 GB | ~6 |
| 12 GB VRAM | Runs, slowly | IQ4_XS (good) | 11.2 GB | 3.3 GB | ~7 |
| 16 GB VRAM | Runs on GPU | IQ4_XS (good) | 14.5 GB | – | ~27 |
| 24 GB VRAM | Runs on GPU | Q6_K (≈ lossless) | 20.5 GB | – | ~32 |
| 32 GB VRAM | Runs on GPU | Q6_K (≈ lossless) | 20.5 GB | – | ~59 |
Ready-made quantizations on Hugging Face: KnutJaegersberg/Tri-21B-Think-gguf · saeha/Tri-21B-AWQ-SRT
Benchmark results
Epoch AI has not published results for Tri-21B yet. New models are usually evaluated within days to weeks.