Slonik-27B-oQ8-mtp

Slonik-27B-oQ8-mtp

Helps write and change PostgreSQL source code with a 27B coding model, 8-bit quantized to run locally on Apple Silicon Macs through oMLX. For PostgreSQL developers who want a local assistant.

Made by kakoka on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: Slonik 27B comes from Postgres Professional and is a coding model. Its training centred on jobs that edit PostgreSQL's own source. This copy is the text-only 8-bit oQ8 quantization for MLX on Macs with Apple Silicon. It keeps the model's native multi-token prediction weights, so oMLX can run it faster with its Lightning MTP setting.

Who it suits: PostgreSQL developers and database engineers who own an Apple Silicon Mac with plenty of memory and want a coding model that runs on their own machine.

Industry | Who uses it | How they use it | Example
Database engineering | PostgreSQL contributors | Ask for patches and tests against the PostgreSQL source | A C fix for a query planner bug
Data platforms | Database administrators | Draft and review SQL and query plans locally | A rewrite of a queue polling query
Software | Mac-based developers | Run a coding model offline on an Apple Silicon Mac | A private coding helper on a workstation
Research | Quantization testers | Compare 8, 6 and 4-bit builds for speed | Median decode speed per build

You can skip it if you do not use an Apple Silicon Mac, or if you need image input, since this is the text-only build.

Where to use it: It runs in oMLX, a local MLX server app, on Macs with Apple Silicon.

• oMLX on an Apple Silicon Mac
• MLX, the Apple machine learning framework the weights are built for
• The tokenizer and chat template included in the download

Use case | Who | What they get back
PostgreSQL code changes | Contributors | Patches and regression tests proposed by the model
SQL and plan work | Database administrators | SQL statements whose results matched the original
Local coding chat | Developers | Fast text answers from a model on your own Mac

How to set it up: You need an Apple Silicon Mac with enough memory for the weights, which take 29.08 GB on disk, and the oMLX app.

• An Apple Silicon Mac with ample unified memory
• The oMLX app
• About 29.08 GB of disk space for the weights

1. Download the repository into a local folder.
2. Add that folder in oMLX.
3. Turn on Lightning MTP for this model.

hf download kakoka/Slonik-27B-oQ8-mtp --local-dir ./Slonik-27B-oQ8-mtp

How to use it: After you add the folder in oMLX and enable Lightning MTP, chat with the model there. The tokenizer and chat template ship with the files, so no extra setup is needed.

Tool | What it does
hf download | Fetches the model into a local folder
oMLX | Loads and serves the model on a Mac
Lightning MTP | Uses the kept multi-token prediction weights to decode faster
Text Only mode | Drops the vision part, which this build lacks

Example: Describe a planner bug in PostgreSQL where an OR condition lets partition pruning skip the DEFAULT partition and lose a row, and ask for a fix. The card reports that the first attempt gave a C fix, and a follow-up added a regression test with its expected output.

For newcomers and experts: Quantizing shrinks a model by storing its numbers with fewer bits, so it fits in less memory. This one is a coding model that knows the PostgreSQL source, squeezed to 8 bits. A first small try is to run the download command and chat with it in oMLX.

For experts, the build stores weights as 8-bit affine values in groups of 64, and oMLX 0.7.4.dev1 did the conversion. The remaining floating-point tensors stay in BF16, and all 15 native MTP tensors are preserved. The trade is memory against speed and accuracy, since the card compares oQ8, oQ6e and oQ4e builds on one Mac Studio.

Start from the BF16 Slonik 27B checkpoint
Quantize to 8-bit in oMLX, text only
Keep the 15 MTP weight tensors
Add the folder to oMLX
Turn on Lightning MTP and chat

Good to know: Every speed figure below is a test with thinking off, run on one top-end Mac Studio. Treat the numbers as a rough guide.

• Test scope: The speed figure is a median of three warmed runs of one prompt, so it is a single data point.
• SQL check: All 72 checks of proposed SQL matched the original results, row order included.
• Bug-fix check: The final patch passed 15 query and plan checks and all 231 core regression tests.

Title: Median decode speed in oMLX
oQ4e with Lightning MTP: 78.98 tok/s
oQ8 with Lightning MTP: 61.88 tok/s
oQ6e with Lightning MTP: 57.71 tok/s
oQ8 without Lightning MTP: 27.92 tok/s

• License: Apache 2.0, for the source checkpoint and this quantized copy.
• Cost to run: none named, the source lists no API key, account or paid service, though it needs a Mac with plenty of memory.
• Data it reaches: only the prompts you give it, plus the files you download.
• Not for: image input, since this is the text-only build, or for machines other than Apple Silicon Macs.

On Hugging Face, 13 November 2026

• Task: text generation
• Library: mlx
• Parameters: 27.3B
• Weights: about 109 GB in U32
• Formats: Safetensors
• Downloads in the last 30 days: 498
• Likes: 3
• License: apache-2.0
• Based on: dglhf/Slonik
• Tags: safetensors, qwen3_5, omlx, slonik, quantized, postgresql, code, text-generation, conversational, 8-bit

Links

• Model: https://huggingface.co/kakoka/Slonik-27B-oQ8-mtp
• Files: https://huggingface.co/kakoka/Slonik-27B-oQ8-mtp/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download kakoka/Slonik-27B-oQ8-mtp

Then follow the usage in its model card. Nothing in this kit runs by itself.