Models & tools
Data as of Oct 13, 2026. Benchmarks are single-machine third-party numbers. Each row links its sources.
Runtimes & servers
| Tool | Engine | APIs | Best for | License | Notes | Src |
|---|---|---|---|---|---|---|
| llama.cpp C++ CLI + server | GGUF / Metal | OpenAI-compatible (llama-server) | Maximum control; prompt caching flags | MIT (open source) | Prompt caching via --cache-prompt, --cache-reuse N and slot save/restore.brew install llama.cpp | [1] |
| LM Studio GUI app + headless llmster | llama.cpp + MLX | OpenAI-compatible | GUI users; model browsing; MLX engine prefill | Closed source (free) | One M4 Pro test (Qwen3.6-35B-A3B): MLX engine prefill ~30s vs ~75s on llama.cpp. 0.4.20 (Jul 2026) added the Bionic agent app.https://lmstudio.ai | [1][2] |
| MLX / mlx-lm Python library + CLI server | MLX | OpenAI-compatible (mlx_lm.server) | Fastest single-stream generation; LoRA/QLoRA fine-tuning | MIT (Apple ml-explore) | In one M3 Ultra test, mlx-lm 0.31.3 hit 107.5 tok/s vs 77.7 for Ollama GGUF (Qwen3.5 9B 4-bit) at 5.2 GB vs 15 GB peak memory.pip install mlx-lm | [1] |
| Ollama CLI + server | llama.cpp (GGUF) + native MLX runner since 0.19 | Ollama API, OpenAI-compatible | Easiest start; `ollama launch claude` for Claude Code | MIT (open source) | MLX runner (`-mlx` tags) since Mar 31 2026. GGUF default loads a 262k context and inflates memory: cap it.https://ollama.com/download | [1][2] |
| oMLX Server + menu-bar app | MLX | OpenAI /v1, Anthropic /v1/messages (:8000) | Long-context coding agents; tiered RAM+SSD KV cache that persists across restarts | Apache-2.0 (open source) | Apple Silicon only, macOS 15+. Continuous batching, multi-model (VLM/OCR/embeddings/rerankers). Fast-moving: ~609 open issues (Oct 2026).brew tap jundot/omlx https://github.com/jundot/omlx && brew install jundot/omlx/omlx | [1][2] |
Models
| Model | Architecture | ~4-bit size | Context | SWE-bench Verified | Suggested RAM | When to pick | Src |
|---|---|---|---|---|---|---|---|
| Gemma 4 (26B / 31B) | see vendor | est. ~14-17 GB (0.5 GB/B rule) | n/a in sources | n/a | 24-32 GB+ | Alternative family mentioned in 2026 Mac roundups | [1] |
| GLM-5.2 | see vendor | n/a in sources | n/a in sources | n/a | check size | Mentioned in 2026 Mac roundups; verify size before downloading | [1] |
| Qwen3.5 9B | Dense | ~5.2 GB peak (MLX) | n/a in sources | n/a | 16 GB | Small-RAM Macs; benchmarked at 107.5 tok/s on mlx-lm (M3 Ultra) | [1] |
| Qwen3.6-27B | Dense | ~16 GB (Q4) | n/a in sources | 77.2 | 24 GB+ | Higher SWE-bench than the MoE but slower per token (all params active) | [1][2] |
| Qwen3.6-35B-A3B | MoE (3B active) | ~20 GB | 262k native | 73.4 | 32 GB+ | Default pick for coding agents: fast because only 3B params are active | [1][2] |
"Suggested RAM" is our estimate: weights + context + macOS headroom. Want your tool listed or featured? See sponsors.