whatmodel.sh / Trinity-Large-Preview
Trinity-Large-Preview
↗ share
+ contribute
Trinity-Large-Preview is an open-weights language model from Arcee AI, released January 2026. Its published size is 398B parameters (13B active per token). Licence: openmdw-1.1.
Catalogue identity notes
Comparison-only. Arena entry 'trinity-large-preview' (text). Maker repo arcee-ai/Trinity-Large-Preview @d8e1267: licence tag other (openmdw-1.1). Parameters: card: 'Total parameters ~398B', 'Active parameters per token ~13B'. Released 2026-01: Arcee blog 'Trinity Large' dated Jan 27, 2026. Lightly post-trained preview checkpoint, separate from Trinity-Large-Thinking. HF licence name openmdw-1.1 (Arcee's docs table says Apache 2.0).
$ how it compares
The LLMs ranked either side of Trinity-Large-Preview, out of 828 with a score, then more from Arcee AI.
| # | model | score | where | ||
|---|---|---|---|---|---|
| 247 |  Gemma 2 9B IT SimPO | 59±1 | local | ||
| 248 | Gemma 2 9B IT DPO (Princeton) | 59±1 | local | ||
| 249 | Trinity-Large-Preview | 59±2 | local | ||
| 250 |  GLM-4.7-Flash | 59±1 | api $0.0000/M +local | ||
| 251 |  EXAONE 4.0 32B | 59±1 | local | ||
| more from Arcee AI | |||||
| 199 | Trinity-Large-Thinking | 63±2 | api $0.39/M +local | ||
$ what it's good at
chat & writing▁▁▁▃▅▄▅█▆▆▇▃ #130 of 393 * · 41% of its score · 1 test
Beats LongCat 2.0 (1.6T-A48B), o3-mini, K-EXAONE (236B-A23B) here, though they rank higher overall. Trails LFM2.5 8B-A1B, LFM2.5 2.6B, Nemotron-Cascade-8B here, though they rank lower overall.
| test | result | vs the others in its run | counts |
|---|---|---|---|
| LMArena (text) | 1378.6 ✎ | ▁▁▁▂▃▃▃▆▆▇█▃ beats 218/379 | 41% |
math█··········█ #2 of 2 * · 6.3% of its score · 1 test
other tests3 not in a task · 51% of its score
| test | result | vs the others in its run | counts |
|---|---|---|---|
| GPQA Diamond lab | 63.3% ✎ | ▁▁▁▁▂▄▅▅███▄ beats 121/310 | 34% |
| MMLU-Pro lab | 75.2% ✎ | ▁▁·▁▁▁▃▂▄▅█▁ beats 103/188 | 17% |
| Vectara hallucination leaderboard (HHEM-2.3, factual consistency rate) | 93.1% ✎ | ▁▂▁▂▅█▂▃▄▃▁▁ beats 71/98 | 0.0% |
$ where to run it
llama.cpp · recommended
version
- Install llama.cpp · Homebrew and winget installs name the commands llama-cli and llama-server instead.macOS / Linux
curl -LsSf https://llama.app/install.sh | shWindowsirm https://llama.app/install.ps1 | iex - Run Trinity-Large-Preview · needs ~247.33 GB · from huggingface.co/unslothchat in the terminal
llama cli -hf unsloth/Trinity-Large-Preview-GGUF:UD-Q4_K_XL - Or start a local server and open the address it prints (http://localhost:9931)
llama serve -hf unsloth/Trinity-Large-Preview-GGUF:UD-Q4_K_XL
LM Studio
version
- Install LM Studio · Open the app once before using the lms command.
- Run Trinity-Large-Preview · needs ~247.33 GB · from huggingface.co/unsloth
- Or download it in the terminal
lms get https://huggingface.co/unsloth/Trinity-Large-Preview-GGUF@UD-Q4_K_XL -y - Then run lms ls to see its name, and lms chat followed by that name
Official weights and documentation: huggingface.co/arcee-ai ↗
$ how the score was worked out · Arcee AI vendor-reported benc… 57%, LMArena leaderboard dataset 41%, Vectara hallucination leaderb… 0%
Each result implies a score on its own, after adjusting for how hard its test is and who ran it. The score is their weighted average, with a small pull toward the middle for thinly tested models. Each test's weight is its counts column above; here it's added up by who ran the tests. How it works
| source | results | share of the score |
|---|---|---|
 Arcee AI vendor-reported benchmark scores for its own Trinity models (Hugging Face model cards) · lab-reported | 3 | 57% |
 LMArena leaderboard dataset | 1 | 41% |
 Vectara hallucination leaderboard (vectara/hallucination-leaderboard repository) | 1 | 0% |
| prior · pull toward the middle | 2% | |
| weighted average | = 58.9 |
$ improve this page
Use a short GitHub form or edit the data through a pull request. Every number needs a source; here's which ones we can use.
data updated · sources and licences · json