It is not positioned as a generally best community artifact. The goal is a cleaner apples-to-apples parity pair derived from the same original checkpoint.

Source

Upstream checkpoint: google/gemma-3-1b-it
Converted from the original Hugging Face checkpoint by mesh-llm

Validation

Validated on studio54.local against the paired GGUF artifact with the mesh-llm exact prompt suite.

Prompt	GGUF f16	MLX bf16
`primary`	`blue`	`blue`
`alt-green`	`green`	`green`
`alt-red`	`red`	`red`
`capital-france`	`Paris`	`Paris`
`primary-colors`	`Red, Green, Blue`	`Red, Green, Blue`
`two-plus-two`	`2 + 2 = 4`	`2 + 2 = 4`
`largest-planet`	`Jupiter`	`Jupiter`
`breathing-gas`	`Oxygen`	`Oxygen`
`opposite-hot`	`Cold`	`Cold`
`banana-color`	`Yellow`	`Yellow`
`after-monday`	`Tuesday`	`Tuesday`

Summary:

GGUF exact: PASS
MLX exact: PASS
parity: MATCH

Files

model.safetensors
model.safetensors.index.json
tokenizer and config files

Notes

This repo exists to support reproducible parity testing in mesh-llm.

meshllm/gemma-3-1b-it-parity-bf16-mlx

Gemma 3 1B IT Parity MLX BF16

Purpose

Source

Validation

Files

Notes

meshllm/
gemma-3-1b-it-parity-bf16-mlx