April 2026 TLDR Setup for Ollama and Gemma 4 26B on a Mac mini

Original Article ↗ Hacker News Discussion ↗

Ollama’s new setup for running Google’s Gemma 4 26B model on Apple Silicon Macs is prompting broader comparisons between local LLM stacks like Ollama, LM Studio, llama.cpp, MLX-based tools, and Unsloth Studio. Commenters debate trade‑offs in openness, ethics, UX, performance, and hardware requirements, with several noting that raw speed and feature support (especially tool calling and multimodal) vary widely by backend, quantization, and platform. Many report that while today’s open‑weight models and local setups are impressive for experimentation and smaller coding or agentic tasks, they still fall short of proprietary models such as Claude Sonnet for complex real‑world development work.

Ollama vs Alternatives (LM Studio, llama.cpp, others)

  • Many argue there’s “no reason” to use Ollama over llama.cpp or LM Studio, citing slower speed, worse defaults (e.g., short context), and lagging model support.
  • Others find Ollama convenient as a backend for various frontends, and appreciate simple commands like ollama pull and OpenAI-compatible APIs.
  • LM Studio is often recommended for beginners as a more polished GUI and model manager, but it’s closed source and desktop-focused.
  • llama.cpp is seen as the “real” core engine: faster, more up to date, and more configurable, though initially perceived as harder to set up.

Licensing, Openness, and Ethics

  • Strong criticism that Ollama started as a derivative of llama.cpp without proper attribution and “quasi-open source” behavior (e.g., mangling GGUF filenames, limited upstream contributions).
  • Counterpoint: the code is MIT-licensed and publicly available, so in a narrow sense it is open source; issues center on attribution and ecosystem behavior, not access.

Performance Comparisons

  • Benchmarks conflict:
    • On an M4 Mac mini, one report finds Ollama ~25% faster than LM Studio for the same Gemma 4 model.
    • On an older AMD GPU, others find llama.cpp fastest, LM Studio 3× faster than Ollama, and Ollama the slowest across several models.
  • Perceived load times also differ; some feel Ollama loads models “nearly instantly” vs llama.cpp.

Usability and UX

  • Ollama is praised for lowering the barrier: easy installs, simple CLI, no Hugging Face account, good default server functionality.
  • Critics say its abstraction hides crucial details (model size, quantization, architecture optimizations), which misleads users about what they’re actually running.
  • Several note that OSS tools often lack strong UX; LM Studio is cited as an exception despite being closed.

Gemma 4 Models: Early Impressions

  • On an M4 Mac mini (24 GB), the ~10 GB Gemma 4 variant is described as fast and usable for small coding tasks; the ~20 GB variant works but is sluggish and RAM-heavy.
  • Multiple reports: Gemma 4 is strong at tool use and data extraction, but early agentic coding performance is underwhelming compared to specialized coding models like Qwen 3.5.
  • Some warn that early negative impressions may come from broken tokenizers or quantizations; implementations are still rapidly evolving.

Tool Calling & Backend Reliability

  • Many tool-calling failures reported across LM Studio, llama.cpp, and Ollama (especially on launch day).
  • Consensus: new open-weight releases are often buggy for weeks; users should expect to update engines and quantizations frequently and file bug reports.
  • Proxies and shims (e.g., “tricks” layers) are used to emulate missing capabilities or fix prompt templates.

Local vs Cloud Models & Expectations vs Claude

  • Several say open models are useful for moderate tasks and privacy-sensitive workflows, but not yet close to Claude Sonnet/Opus or top proprietary coders for complex projects.
  • Advice for people considering hardware purchases: first try hosted versions (e.g., via aggregators) to understand capabilities and limitations.
  • Some users experiment with open models for “agentic” coding but still fall back to Claude or other proprietary models for hard problems.

Hardware and Engine Notes (Macs, GPUs, MLX)

  • Users report running Gemma 4 26B on MacBook Air/mini with 24–32 GB RAM; throughput in the ~20–40 tok/s range is seen as borderline for agents but fine for chat.
  • Unified memory on Apple Silicon avoids separate CPU/GPU VRAM constraints but doesn’t eliminate latency concerns for large models.
  • MLX/oMLX support for Gemma 4 is emerging; current builds support basic chat, with partial or in-progress support for special tokens and tool calling.