NotHereNorThere/Coral-2-4b
4B • Updated • 21
Merged models to combine strengths without retraining or heavy fine tuning.
Note the full first release of Coral 2. by far the best model so far in terms of vibes. no CoT, period.
Note incremental experiment. decent/on par with 1.5 0.6b.
Note overall genuinely capable for edge inference on medium-simple tasks.
Note Huge leap up from Coral 1B despite half the parameters. More base models and more data.
Note First ever, mediocre at best. Use for whatever you'd use Llama3.2 1b uncensored for.