Pixtral-12B-GGUF Modelfile (Q4_K_M)
---------------------------------
Tested with: Ollama v0.11.X-->v0.12.6(latest)
Quantization: Q4_K_M (Quant created by = ggml-org)
Quality: Very-Good (Updated 2025/10/28)
Real-world usability: Recommended!
----------------------------------------------------
Vision Notes:
Some users may need to set the context value -or- "num_ctx"
value to ~9K-->19K.
Personally tested with: num_ctx=9982 and num_ctx=19982
-----------------------------------------------------------
Created by:
EnlistedGhost (aka Jon Zaretsky)
Original GGUF by: https://huggingface.co/ggml-org
Original GGUF type: Static Quantize (non-iMatrix)
----------------------------------------------------------
| Warning! - iMatrix Quantize seems to suffer in regards |
| to vision quality, but is still made available |
----------------------------------------------------------
Goal:
To provide the FIRST actually functional and usable
GGUF model version of the Mistral Pixtral-12B for
direct-usage with Ollama!
Currently, there are NO USABLE OR WORKING versions
of this model that are usable with Ollama...
---------------------------------------------------
Big/Giant/Huge Thank You:
(ggml-org, bartowski, and the Ollama team)
ggml-org: Working mmproj-pixtral vision projector!
Bartowki: Working I-Matrix Quants that can be paired with ggml-org vision projector!
Ollama team: Because without them, this wouldn't be possible in the first place!
------------------------------------------------------------------------------------
Import our GGUF quant files:
(Assuming: Linux Operating System)
(Assuming: downloaded files are stored in "Downloads" directory/folder)
FROM ~/Downloads/mmproj-pixtral-12b-f16.gguf FROM ~/Downloads/Pixtral-12B-Q4_K_M.gguf
------------------------------------------------------------------------
Set Default System-Message/Prompt:
SYSTEM """
!!!-WARNING-!!!
(Do not modify for: "recommended" configuration and behavior)
!!!-OPTIONAL-!!!
Pixtral-12B by default does NOT include a system-prompt, however -
you can choose to input one within this section of the Ollama-Modelfile.
Please be aware that you can possibly damage the linking -
between the Pixtral-->VisionProjector within the system-prompt field.
BE CAREFUL!
"""
-------------------------------------------------------------------
Define model-chat template (Thank you to: @rick-github for this mic-drop)
Link to @rick-github post: https://github.com/ollama/ollama/issues/6748#issuecomment-3368146231
TEMPLATE [INST] {{ if .System }}{{ .System }} {{ end }}{{ .Prompt }} [/INST]
Below are stop params (required for proper "assistant-->user" multi-turn)
PARAMETER stop [INST] PARAMETER stop [/INST]