Update README.md
Browse files
README.md
CHANGED
|
@@ -10,8 +10,29 @@ tags:
|
|
| 10 |
license: apache-2.0
|
| 11 |
---
|
| 12 |
|
|
|
|
|
|
|
| 13 |
<h2>Gemma-3-1b-it-MAX-NEO-Imatrix-GGUF</h2>
|
| 14 |
|
| 15 |
Google's newest Gemma-3 model with Neo Imatrix and Maxed out quants.
|
| 16 |
|
| 17 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 10 |
license: apache-2.0
|
| 11 |
---
|
| 12 |
|
| 13 |
+
(quants uploading, examples to be added)
|
| 14 |
+
|
| 15 |
<h2>Gemma-3-1b-it-MAX-NEO-Imatrix-GGUF</h2>
|
| 16 |
|
| 17 |
Google's newest Gemma-3 model with Neo Imatrix and Maxed out quants.
|
| 18 |
|
| 19 |
+
Recommend quants IQ3s / IQ4XS / Q4s for best results for creative.
|
| 20 |
+
|
| 21 |
+
Recommend q5s/q6/q8 for general usage.
|
| 22 |
+
|
| 23 |
+
Q8 is a maxed quant only, as imatrix has no effect on this quant.
|
| 24 |
+
|
| 25 |
+
Note that IQ1 performance is low, whereas IQ2s are passable.
|
| 26 |
+
|
| 27 |
+
"MAXED"
|
| 28 |
+
|
| 29 |
+
This means the embed and output tensor are set at "BF16" (full precision) for all quants.
|
| 30 |
+
|
| 31 |
+
This enhances quality, depth and general performance at the cost of a slightly larger quant.
|
| 32 |
+
|
| 33 |
+
"NEO IMATRIX"
|
| 34 |
+
|
| 35 |
+
A strong, in house built, imatrix dataset built by David_AU.
|
| 36 |
+
|
| 37 |
+
- more to follow -
|
| 38 |
+
|