FrankenGemma4 MLX 4Bit is the polished local release of the FrankenGemma4 line. The current public lead branch is Frankengemma4 V1, and that is the branch I recommend for normal use on Apple Silicon.
Parameter note: E4B-family model: approximately 4.5B effective parameters / ~8B including embeddings under Google's Gemma 4 convention. This MLX 4-bit export's on-disk tensor count is smaller because it is quantized.
This repo is intended to be the public MLX 4-bit release:
stamsam/FrankenGemma4_MLX_4Bit
recommended default artifact: the fused MLX Q4 checkpoint
current lead branch: Frankengemma4 V1
What This Release Is
This release comes from a two-stage lineage:
Original frankenmerge
Passthrough layer-stacking between the reasoning donor and the coding donor.
Co-base repair line
A symmetric linear merge across the shared language stack.
Followed by targeted MLX LoRA repair passes for seam control, leak suppression, structured chat, coding repair, and daily chat.
Recommended Default
Use the root MLX Q4 checkpoint as the default download for this repo.
Local Benchmark Snapshot
These are local custom evals from the development workflow.
The detailed benchmark artifacts for this release live in the benchmarks/ folder.
Q4 Snapshot
Metric
Score
Exact Overall
68.75
Reasoning
71.43
JSON
85.71
Code
71.43
Integration
54.55
OpenClaw / Hermes / Agentic Snapshot
Model
Coding
Daily Chat
Structured Chat
Tool Use
Agentic
Total
FrankenGemma4 V1
4
8
7
9
10
38
FrankenGemma4 Structured-1600
4
6
7
9
10
36
FrankenGemma4
2
10
4
9
10
35
SuperGemma4 E4B Ablit
2
8
7
8
10
35
Google Gemma 4 E4B IT
2
8
6
9
10
35
Reasoning Donor
2
8
4
0
10
24
Lead Branch Retention
Model
Security Defense
Blunt Critique
Uncensored Creative
Abliteration Meta
Profane Rewrite + Note
Prompt Injection Defense
Total
SuperGemma4 E4B Ablit
9
8
9
7
10
7
50
Frankengemma4 V1
10
7
9
7
10
6
49
Current Strengths
Good local MLX/Q4 behavior on Apple Silicon
Stronger tool discipline than the reasoning donor
Better structured output after the repair passes
Retains the ablation-style behavior better than the raw franken line while staying close to the donor parents
Current Caveats
Some prompts still show thought leakage
This is still a local benchmark story, not a broad held-out public leaderboard claim
The retention check shows the model keeps most of the ablation behavior, but not quite as much as the dedicated ablated donor
