github-starred/ollama
2
0
Fork 0
mirror of https://github.com/ollama/ollama.git synced 2026-05-07 00:22:43 -05:00
Code Issues 14.5k Packages Projects Releases 120 Wiki Activity

[GH-ISSUE #1543] Better model quantization defaults from ollama.com #62879

New Issue
Closed
opened 2026-05-03 10:35:41 -05:00 by GiteaMirror · 7 comments
GiteaMirror commented 2026-05-03 10:35:41 -05:00
Owner
Copy Link

Originally created by @knoopx on GitHub (Dec 15, 2023).
Original GitHub issue: https://github.com/ollama/ollama/issues/1543

Is there a reason latest tag on the model hub points, by default, to the older q4_0 quants? The newer k_m/s are
supposedly better and the size difference is usually just a few hundred megabytes, it would be nice if it defaulted to those instead.

Originally created by @knoopx on GitHub (Dec 15, 2023). Original GitHub issue: https://github.com/ollama/ollama/issues/1543 Is there a reason `latest` tag on the model hub points, by default, to the older `q4_0` quants? The newer `k_m/s` are supposedly better and the size difference is usually just a few hundred megabytes, it would be nice if it defaulted to those instead.
GiteaMirror added the ollama.com label 2026-05-03 10:35:41 -05:00
GiteaMirror commented 2026-05-03 10:35:44 -05:00
Author
Owner
Copy Link

@choltha commented on GitHub (Dec 22, 2023):

Difference in perplexity (can be used to measure model response quality) is listed here:
https://github.com/ggerganov/llama.cpp/discussions/2094#discussioncomment-6351796
Relevant lines:

Q4_0 : 3.50G, +0.2499 ppl @ 7B
Q3_K_L : 3.35G, +0.1803 ppl @ 7B
Q4_K_S : 3.56G, +0.1149 ppl @ 7B
Q4_K_M : 3.80G, +0.0535 ppl @ 7B
F16 : 13.00G @ 7B

F16 is the baseline reference for perplexity

Surprisingly even Q3_K_L is listed with lower perplexity than Q4_0 while being smaller. This is also the recommendation by "the bloke":

Q4_0: legacy; small, very high quality loss - prefer using Q3_K_M

I would support to switch to Q4_K_M by default (best compromise)

If not wanted because of size-increase, at least switch to Q3_K_M which would be better by any measure.

<!-- gh-comment-id:1868047871 --> @choltha commented on GitHub (Dec 22, 2023): Difference in perplexity (can be used to measure model response quality) is listed here: https://github.com/ggerganov/llama.cpp/discussions/2094#discussioncomment-6351796 Relevant lines: > Q4_0 : 3.50G, +0.2499 ppl @ 7B > Q3_K_L : 3.35G, +0.1803 ppl @ 7B > Q4_K_S : 3.56G, +0.1149 ppl @ 7B > Q4_K_M : 3.80G, +0.0535 ppl @ 7B > F16 : 13.00G @ 7B F16 is the baseline reference for perplexity Surprisingly even Q3_K_L is listed with lower perplexity than Q4_0 while being smaller. This is also the recommendation by "the bloke": > Q4_0: legacy; small, very high quality loss - prefer using Q3_K_M I would support to switch to Q4_K_M by default (best compromise) If not wanted because of size-increase, at least switch to Q3_K_M which would be better by any measure.
GiteaMirror commented 2026-05-03 10:35:44 -05:00
Author
Owner
Copy Link

@choltha commented on GitHub (Dec 22, 2023):

I would create a PR if someone points me to the place where the defaults for this are set.

<!-- gh-comment-id:1868052626 --> @choltha commented on GitHub (Dec 22, 2023): I would create a PR if someone points me to the place where the defaults for this are set.
GiteaMirror commented 2026-05-03 10:35:45 -05:00
Author
Owner
Copy Link

@mchiang0610 commented on GitHub (Mar 11, 2024):

Thank you for sharing this. We will be working to have more sane defaults from ollama.com; sorry about this.

<!-- gh-comment-id:1989199393 --> @mchiang0610 commented on GitHub (Mar 11, 2024): Thank you for sharing this. We will be working to have more sane defaults from ollama.com; sorry about this.
GiteaMirror commented 2026-05-03 10:35:46 -05:00
Author
Owner
Copy Link

@choltha commented on GitHub (Mar 12, 2024):

Thanks for your work!

<!-- gh-comment-id:1991021091 --> @choltha commented on GitHub (Mar 12, 2024): Thanks for your work!
GiteaMirror commented 2026-05-03 10:35:47 -05:00
Author
Owner
Copy Link

@jtsorlinis commented on GitHub (Jun 14, 2024):

Any updates on this? Would be great to default to something like q4_k_m

<!-- gh-comment-id:2167499436 --> @jtsorlinis commented on GitHub (Jun 14, 2024): Any updates on this? Would be great to default to something like q4_k_m
GiteaMirror commented 2026-05-03 10:35:48 -05:00
Author
Owner
Copy Link

@DuckyBlender commented on GitHub (Jul 16, 2024):

related to #5425

<!-- gh-comment-id:2231874313 --> @DuckyBlender commented on GitHub (Jul 16, 2024): related to #5425
GiteaMirror commented 2026-05-03 10:35:48 -05:00
Author
Owner
Copy Link

@jmorganca commented on GitHub (Dec 29, 2024):

The default is now q4_k_m - and in the future additional, higher-quality 4-bit quantization formats are looking to be supported (e.g. AWQ)

<!-- gh-comment-id:2564817498 --> @jmorganca commented on GitHub (Dec 29, 2024): The default is now q4_k_m - and in the future additional, higher-quality 4-bit quantization formats are looking to be supported (e.g. AWQ)
Sign in to join this conversation.
No Branch/Tag Specified
Branches Tags
Labels
Clear labels
amd
api
app
bug
build
cli
cloud
compatibility
context-length
create
docker
documentation
embeddings
feature request
feedback wanted
good first issue
gpt-oss
gpu
harmony
help wanted
image
install
intel
js
launch
linux
macos
memory
mlx
model
needs more info
networking
nvidia
ollama.com
performance
pull-request
Mirrored from GitHub Pull Request
 python
question
registry
rendering
thinking
tools
top
vulkan
windows
wsl
No Label ollama.com
Milestone
No items
No Milestone
Projects
Clear projects
No project
Assignees
Clear assignees
GiteaMirror ninjasurge
No Assignees
1 Participants
Notifications
Due Date
No due date set.
Dependencies

No dependencies set.

Reference: github-starred/ollama#62879
Delete Branch "%!s()"

Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?