kyr0/Gemma-4-Waldwicht-Sproessling

Sprössling - MLX Edition - 3.17 GB

Overview

Waldwicht-Sproessling is a high-quality 5-bit mixed-precision quantization of Gemma 4 E2B (2.3B effective parameters). At 3.17 GB it matches or exceeds uniform 4-bit (3.22 GB) on all quality metrics while being slightly smaller - discovered through a 50+ configuration ablation study on Apple Silicon.

This is the balanced variant: attention and MLP both stay at 5-bit for maximum reasoning fidelity, while only PLE/embeddings are reduced to 3-4 bit. Choose this over Winzling when you need the best possible quality within a ~3 GB budget.

Part of the Waldwicht family: Winzling (2.96 GB) - Sproessling (3.17 GB) - Juengling (3.86 GB)

Quick Start

For a quick test, you can install the a current version of mlx-lm from PyPI and run mlx_lm.generate with the model name and a prompt:

pip install mlx-lm
python -m mlx_lm.generate --model kyr0/Gemma-4-Waldwicht-Sproessling --prompt "Explain quantum tunneling in simple terms."

However, we recommend using the Waldwicht Inference Server.

Waldwicht Inference Server

Together with this model release, we are also launching the Waldwicht Inference Server, a high-performance server optimized for running our quantized models on Apple Silicon. The server is open-source and available on GitHub: Waldwicht Inference Server

Quantization Specification

Component Group	Weight Key Patterns	Bits	Group Size
Attention	q_proj, k_proj, v_proj, o_proj	5	64
MLP	gate_proj, up_proj, down_proj	5	64
PLE embeddings	embed_tokens_per_layer	3	64
PLE gate/projection	per_layer_input_gate, per_layer_projection	4	64
Main embed + LM head	embed_tokens, lm_head	3	64
Norms, scalars	(not quantized)	BF16	-

Key Metrics

Metric	Value
Model size	3.17 GB
Memory footprint	2.83 GB peak (MacBook Air M4 24 GB)
Compression from BF16	3.0x
Throughput	48.6 tok/s avg (MacBook Air M4 24 GB)
Correctness (20-prompt)	9.25 / 10.0
Completion (20-prompt)	9.15 / 10.0
Reasoning hygiene (20-prompt)	9.35 / 10.0
Expanded AQ gate	[OK] PASS (threshold >= 7.0)

Benchmark Results

Evaluated on a 20-prompt diverse benchmark covering 6 task categories: function calling (3), multi-step function calling (1), code generation (4), translation (4), multi-step reasoning (4), and creative writing (4). Each response scored on correctness, completion, and reasoning hygiene (0-10 scale).

Overall: 9.25 / 9.15 / 9.35 (correctness / completion / reasoning hygiene)

Comparison with Other Variants

Config	Size	RAM	tok/s	Corr.	Comp.	Reas.	Gate
Winzling (5-bit mixed B)	2.96 GB	2.63 GB	51.5	8.75	8.95	9.05	PASS
Sproessling (this model)	3.17 GB	2.83 GB	48.6	9.25	9.15	9.35	PASS
Juengling (uniform 5-bit)	3.86 GB	3.52 GB	47.4	9.40	9.35	9.50	PASS

How This Config Was Discovered

Sproessling applies the relative weighting transfer pattern to a 5-bit base (Section 7.9, Variant A). The component sensitivity ordering - attention > MLP > gate > PLE/embed - was discovered through 28 component-level experiments at 4-bit base, then transferred to 5-bit:

• Attention and MLP remain at 5-bit (full base precision) - these carry the reasoning and code generation capability
• PLE embeddings and main embed/LM-head drop to 3-bit (-2 offset) - these are lookup tables tolerant of compression
• PLE gate/projection drops to 4-bit (-1 offset)

The result: 5-bit precision on the core compute path (attention + MLP) delivers quality that matches or exceeds uniform 4-bit across all 6 task categories, while 3-bit PLE/embeddings keep the total size 0.05 GB below uniform 4-bit.

Key Advantages over Winzling (2.96 GB)

• Higher overall quality: +0.50 correctness, +0.20 completion, +0.30 reasoning hygiene
• 5-bit MLP (vs 4-bit in Winzling) preserves reasoning and code generation fidelity
• Comparable throughput (48.6 vs 51.5 tok/s)

Research Paper

Full details: Pushing Below 3 GB: Component-Level Mixed-Precision Quantization for Gemma 4 E2B on Apple Silicon (Homberg, April 2026). 50+ configurations, 4 experimental phases, 20-prompt diverse benchmark with LLM-judge scoring.

Base model

For the base model documentation, see Gemma 4 E2B - the smallest dense model in the Gemma 4 family, with 2.3B effective parameters and a 128K token context window.