Local AI Infrastructure Notes (9/15) — Replacing the Local LLM: From Qwen3.5 to Gemma 4

A generational swap between MoE models. Speed, memory, accuracy, context — all improved.

Summary

  • Replaced Qwen3.5-35B-A3B-4bit with gemma-4-26b-a4b-it-4bit on Mac Mini M4 32GB
  • Speed +7%, RAM -4GB, accuracy 94%→100%, context 4× (32K→131K), multimodal support added
  • All existing pipeline tests passed: recall-tree, retain-merge, micro-cycle, and more

Background: Why a Local LLM

Cloud LLMs are powerful but not cost-effective for all workloads. Memory management, classification, and summarization tasks that run automatically hundreds of times per session will produce runaway token costs when routed through a cloud provider.

This was directly observed during a migration attempt that connected a multi-agent orchestration platform's memory pipeline to a cloud LLM — token consumption became uncontrollable. The architecture was restructured to isolate a local model for mechanical tasks. The local LLM handles classification, summarization, and tag extraction; the cloud LLM handles tasks requiring strategic judgment.

Qwen3.5-35B-A3B (Alibaba) had been running on this layer for an extended period. When Google released Gemma 4 — another MoE architecture — a same-class generational replacement became viable.

Body

1. Model Spec Comparison

Item Qwen3.5-35B-A3B (Alibaba) Gemma-4-26B-A4B (Google)
Total parameters 35B 26B
Active parameters 3B 3.8B
Architecture MoE (35B/3B) MoE (26B/3.8B, 128 experts, top-8)
Quantization 4-bit GGUF 4-bit MLX safetensors
Disk 19 GB 15 GB
Context 32K 131K (256K supported)
Vision None Multimodal supported

Qwen has more total parameters, but Gemma has more active parameters (3B vs 3.8B). In MoE models, active parameters determine actual inference behavior — Gemma has the edge here.

Disk footprint is also smaller: 15 GB vs 19 GB. The MLX safetensors format runs natively on Apple Silicon.

Considerations when selecting a local LLM on Apple Silicon (Mac Mini M4): - Unified memory architecture means model size maps directly to RAM consumption — no VRAM/RAM split. - MLX leverages the Apple Neural Engine; it is more efficient than GGUF on the same hardware. - A 15 GB model on a 32 GB system leaves sufficient headroom for OS and services. 19 GB is tight.

2. Speed Benchmarks

Measured on Mac Mini M4 32GB, oMLX 0.3.4.

Item Qwen3.5 Gemma 4 Delta
Generation speed (256 tokens, 3-run avg) 27.3 tok/s 29.1 tok/s +7%
oMLX stats average 14.2 tok/s (long-term cumulative) 22.5 tok/s (early measurement) +58%
RAM usage ~19 GB ~15 GB -4 GB
Free RAM ~13 GB ~17 GB +4 GB

The 256-token controlled benchmark shows a 7% improvement. The oMLX stats comparison shows 58%, but this compares Qwen's long-term cumulative average (including intermittent stalls) against Gemma's early measurement set — a direct comparison is not statistically fair. Long-term recalibration is required after extended operation.

The RAM savings are the more immediately actionable result. Freeing 4 GB allows reallocation to other services: dashboard, monitoring, and embedding server.

3. Accuracy Tests

Ran the T01–T14 test suite used by OpenClaw's memory-runner pipeline.

Test Qwen3.5 Gemma 4
T01–T10 (basic instruction / file ops) 10/10 10/10
T13 (multi-tag extraction) 3/3 3/3
T14 (conditional branching) 2/3 3/3
Total 15/16 (94%) 16/16 (100%)

The critical delta is T14: conditional branching. This was Qwen3.5's single failure point. Gemma 4 passes it cleanly. Conditional branching governs "where does this memory get stored" in the memory management layer — a 100% pass rate here is operationally significant.

4. Pipeline Compatibility

Model-agnostic pipeline design was the precondition for this swap. Pipelines that depend on model identity rather than output format cannot be cleanly migrated.

All pipelines passed: - recall-tree: full query test suite passed - retain-merge: TAG_ROUTING keyword classification intact - micro-cycle: lightweight distillation intact - confidence-decay: decay calculation intact - bank-lint: knowledge file integrity check intact - proactive-select: briefing selection logic intact - Embedding (bge-m3 1024d): hybrid search intact

Why the swap had zero pipeline impact: the pipelines depend on output format (JSON, tags), not on model identity. This decoupled design is what makes model replacement safe.

5. What 4× Context Expansion Actually Means

Context expanded from 32K to 131K (up to 256K supported). Practical implications:

  • memoryFlush: longer session summaries can be processed in a single pass
  • retain-merge: more context available for tag classification decisions
  • Reflect distillation: a full day's memory can be processed in one pass without chunking

Longer context does increase per-token processing time, so simply maximizing context length is not the correct approach. The value is headroom when it is needed.

6. Local LLM vs Cloud LLM — Selection Criteria

Local LLM is preferable when: - High-frequency repetitive calls: classification, tag extraction, summarization in automated pipelines running hundreds to thousands of times. Cloud LLM cost scales linearly. - Latency tolerance: 27–29 tok/s is insufficient for real-time conversation but adequate for background batch processing. - Privacy: data does not leave the local machine.

Cloud LLM is preferable when: - Complex reasoning: strategic judgment, multi-step planning, creative generation. - Current information: local models have a training cutoff. - Long context + high throughput: cloud infrastructure has no hardware ceiling.

Practical approach: route mechanical, repetitive tasks to local; route judgment tasks to cloud. The two-tier architecture optimizes both cost and quality simultaneously.

Lessons Learned

oMLX stats sampling bias: Qwen's long-term cumulative average vs Gemma's early sample is not a controlled comparison. Recalibration after extended operation is necessary. Early benchmark results warrant measured interpretation.

Quantization format difference: GGUF (Qwen) and MLX safetensors (Gemma) use different runtimes. On Mac Mini, MLX is native — this advantages Gemma. On Linux/CUDA environments, results may differ.

Conclusion

Gemma 4 is a clean generational replacement: every benchmark metric improved while full pipeline compatibility was maintained. The T14 conditional branching 100% pass rate and the 4 GB RAM reduction are the two results with the most direct operational value.

For developers running local LLMs, two design recommendations:

  1. Run pipeline compatibility tests before committing to a model swap.
  2. Design pipelines to depend on output format (JSON/tags) rather than on model identity — this reduces replacement cost to near zero.

Series overview: Series index

๋Œ“๊ธ€

๋Œ“๊ธ€ ์“ฐ๊ธฐ

์ด ๋ธ”๋กœ๊ทธ์˜ ์ธ๊ธฐ ๊ฒŒ์‹œ๋ฌผ

Agent Memory Engine (2/10) — Building an AI Agent Memory System with SQLite Alone

← 1/10 SQLite Memory Engine f… ๐Ÿ“š Series Index 3/10 memcore Completed → A memory engine that runs without daemons, without dependencies, anywhere ํ•ต์‹ฌ ์š”์•ฝ This post covers how to design an AI agent memory engine (memcore) on a single-file SQLite backend. - Structure: 25 modules, ~6,464 lines, 22 DB tables - Principles: no daemon, single-file portability, no required external dependencies, host-agnostic - Core techniques: 4-layer hybrid search (topic · vector · FTS5 · LIKE), U-tag dialectic-based confidence evolution, ingestion-layer bias rejection rules What This Post Covers One approach to building AI agent memory without a server or cloud vector DB. Specifically: (1) the limitations of text-file-based memory, (2) vector DB alternatives compared and selection criteria, (3) how to layer search tiers, (4) how to evolve confidence scores over time, and (5) how to block biased memory accumulation at the ingestion stage. Design Principles: Why Single-File SQLite Text-file-based ...
Read more »

"ML Foundations (9/9) — PyTorch vs TensorFlow, and the Road to Local LLMs"

์ด๋ฏธ์ง€
← 8/9 Deep Learning Architec… ๐Ÿ“š Series Index (series end) The final part. So far we've focused on models. Now we focus on the tools that actually run them . We'll lay out the philosophical difference between PyTorch and TensorFlow, and trace how Hugging Face transformers , llama.cpp , MLX , and Ollama built the bridge to running large language models on your own machine. By the end you should have the full mental model of "download a pretrained LLM and serve it locally." 0. Learning Objectives Explain the eager-vs-graph difference between PyTorch and TensorFlow. Explain, in graph terms, how autograd automates the backward pass. Use Hugging Face transformers ' AutoModel , AutoTokenizer , and pipeline abstractions. Describe llama.cpp's GGUF quantization, INT4 inference, and CPU-first flow. Describe Apple MLX's use of unified memory and how it differs from PyTorch. Run a local LLM with Ollama and call it through an OpenAI-compatible API. ...
Read more »

"RAG Core Study (14/26) — Evaluation Sets with RAGAS & DeepEval"

์ด๋ฏธ์ง€
Series overview: Series index ๐Ÿ“š Series Index Without an evaluation set, every RAG improvement is a story, not evidence. RAG systems fail in more than one place: retrieval, context selection, answer generation, and grounding. That means "looks better" is not a metric. Part 14 explains how to build a golden dataset , how RAGAS and DeepEval fit into the workflow, and why teams must separate retrieval quality from answer quality if they want tuning decisions to hold up. 0. Prerequisites Part 13 reranking — system quality now depends on multi-stage ranking. Part 1 grounding — a correct-looking answer is not enough. Part 12 Hybrid — multiple retrieval settings need comparison on fixed data. 1. Learning Objectives Build a small but useful golden dataset for RAG. Distinguish retrieval metrics from answer metrics . Use RAGAS and DeepEval in the right roles. Avoid the common traps in judge-model-based evaluation. 2. ํ•ต์‹ฌ ์š”์•ฝ An evaluation set for RAG is ...
Read more »

"ML Foundations (8/9) — Deep Learning Architectures: CNN, RNN, Attention"

์ด๋ฏธ์ง€
← 7/9 Deep Learning Training ๐Ÿ“š Series Index 9/9 PyTorch vs TensorFlow → What did we keep adding on top of an MLP? Vision went CNN, sequences went RNN/LSTM, and eventually both converged on Attention and the Transformer. Each architecture is easier to remember if you read it as what it gave up to gain something else . This part is the one-line summary of the decisive inflection points: LeNet → AlexNet → ResNet → LSTM → Attention → Transformer. 0. Learning Objectives Explain why CNN's convolution, pooling, and weight sharing match images so well. Trace what got unblocked from LeNet → AlexNet → VGG → ResNet as depth grew. State BPTT for RNNs and the vanishing/exploding gradient problem. Write the LSTM gate equations (forget, input, output) with intuition. Write the attention formula in query/key/value form and the scaled dot-product variant. State the precise reasons Transformers replaced RNNs (parallelism + long-range dependencies). 1. ํ•ต์‹ฌ ์š”์•ฝ CNN : weight sharin...
Read more »

"ML Foundations (7/9) — Deep Learning Training: Optimizers, Regularization, Initialization"

← 6/9 Neural Networks ๐Ÿ“š Series Index 8/9 Deep Learning Architec… → Building an MLP, as we did in Part 6, is not the same as getting it to train well . The right optimizer, regularization, initialization, and learning rate — when these line up, deep networks converge. When they don't, the network refuses to learn at all. This part is the catalog of those crafts, with formulas, paper citations, and code in one place. 0. Learning Objectives Compare and write the update rules for SGD, Momentum, Nesterov, and Adam. Explain how Dropout, BatchNorm, and LayerNorm work and where they belong in a model. Derive the variance formulas for Xavier and He initialization and match them to activations. Implement step, cosine, and warmup learning-rate schedules in PyTorch. Explain why gradient clipping is effectively required for RNNs and Transformers. Diagnose the most common training failures (NaN, plateau, overfitting) and apply first-line fixes. 1. ํ•ต์‹ฌ ์š”์•ฝ SGD : \(w \leftarro...
Read more »

OpenClaw to Hermes Migration (2/13) — What to Preserve, Partially Port, or Discard

← 1/13 Current Structure Inve… ๐Ÿ“š Series Index 3/13 What Moves to Hermes → A code review record classifying operational scripts into three tiers: preserve, partial-preserve, and discard. What This Post Covers An asset classification framework (preserve / partial-preserve / discard) for migrating a legacy agent system to a successor A concrete case applying the Strangler Fig + Subset Migration strategy to a memory pipeline Classification results evaluated across LOC, external dependencies, and replaceability axes A method for slimming down a bloated single file (1,310 LOC) using the migration as a natural entry point Problem Definition: Port Everything or Cut Deep? The OpenClaw inventory comprises multiple agents, multiple scheduler/daemon processes, and multiple operational scripts. When moving to the successor system, Hermes, the choice is not binary. A full port carries accumulated technical debt along with it; a minimal port discards hard-won domain logic. Code-revie...
Read more »

AI Agents I Built (5/7) — Building an Automated Blogger API Publishing System

← 4/7 My Life Organizer Agent ๐Ÿ“š Series Index 6/7 Dual → Designing a Markdown-Based Auto-Publish Pipeline with Blogger API v3 + OAuth 2.0 Key Summary Blogger API v3 + OAuth 2.0 enables a pipeline that converts Markdown files into publishable HTML posts blogger_publish.py handles the full flow — frontmatter parsing → body cleanup → HTML conversion → CSS injection → API post — in a single script For bulk publishing, quota management (delay, retry, TOC suppression) is the critical design concern Platform Selection Criteria API access is mandatory for AI agent-driven content publishing. A comparison of platforms: Platform API Cost AdSense Decision WordPress.com REST API Paid plan required Paid only Cost overhead Ghost REST API Self-hosted or paid Manual setup Ops overhead Blogger v3 API Free Native integration Adopted Blogger is free, exposes a REST API, and integrates natively with AdSense. For personal blog automation, this combination i...
Read more »