Run Local LLMs on Mac

Run Local LLMs on Mac

This guide walks you through running a local LLM server on macOS with an OpenAI-compatible API. You get full privacy, zero API costs, and surprisingly good performance on Apple Silicon.

We cover two backends:

Backend	Install	Best at	Format
llama.cpp	brew install llama.cpp	Fastest time-to-first-token, quantized KV cache for low memory	GGUF
omlx	omlx.ai	Fastest token generation, native Metal optimization	MLX (safetensors)

Both expose an OpenAI-compatible /v1/chat/completions endpoint. Hermes works with either one — just point it at http://localhost:8080 or http://localhost:8000.

:::info Apple Silicon only This guide targets Macs with Apple Silicon (M1 and later). Intel Macs will work with llama.cpp but without GPU acceleration — expect significantly slower performance. :::

---

Choosing a model

For getting started, we recommend Qwen3.5-9B — it’s a strong reasoning model that fits comfortably in 8GB+ of unified memory with quantization.

Variant	Size on disk	RAM needed (128K context)	Backend
Qwen3.5-9B-Q4_K_M (GGUF)	5.3 GB	~10 GB with quantized KV cache	llama.cpp
Qwen3.5-9B-mlx-lm-mxfp4 (MLX)	~5 GB	~12 GB	omlx

Memory rule of thumb: model size + KV cache. A 9B Q4 model is ~5 GB. The KV cache at 128K context with Q4 quantization adds ~4-5 GB. With default (f16) KV cache, that balloons to ~16 GB. The quantized KV cache flags in llama.cpp are the key trick for memory-constrained systems.

For larger models (27B, 35B), you’ll need 32 GB+ of unified memory. The 9B is the sweet spot for 8-16 GB machines.

---

Option A: llama.cpp

llama.cpp is the most portable local LLM runtime. On macOS it uses Metal for GPU acceleration out of the box.

Install

brew install llama.cpp

This gives you the llama-server command globally.

Download the model

You need a GGUF-format model. The easiest source is Hugging Face via the huggingface-cli:

brew install huggingface-cli

Then download:

huggingface-cli download unsloth/Qwen3.5-9B-GGUF Qwen3.5-9B-Q4_K_M.gguf --local-dir ~/models

:::tip Gated models Some models on Hugging Face require authentication. Run huggingface-cli login first if you get a 401 or 404 error. :::

Start the server

llama-server -m ~/models/Qwen3.5-9B-Q4_K_M.gguf \
  -ngl 99 \
  -c 131072 \
  -np 1 \
  -fa on \
  --cache-type-k q4_0 \
  --cache-type-v q4_0 \
  --host 0.0.0.0

Here’s what each flag does:

Flag	Purpose
-ngl 99	Offload all layers to GPU (Metal). Use a high number to ensure nothing stays on CPU.
-c 131072	Context window size (128K tokens). Reduce this if you’re low on memory.
-np 1	Number of parallel slots. Keep at 1 for single-user use — more slots split your memory budget.
-fa on	Flash attention. Reduces memory usage and speeds up long-context inference.
--cache-type-k q4_0	Quantize the key cache to 4-bit. This is the big memory saver.
--cache-type-v q4_0	Quantize the value cache to 4-bit. Together with the above, this cuts KV cache memory by ~75% vs f16.
--host 0.0.0.0	Listen on all interfaces. Use 127.0.0.1 if you don’t need network access.

The server is ready when you see:

main: server is listening on http://0.0.0.0:8080
srv  update_slots: all slots are idle

Memory optimization for constrained systems

The --cache-type-k q4_0 --cache-type-v q4_0 flags are the most important optimization for systems with limited memory. Here’s the impact at 128K context:

KV cache type	KV cache memory (128K ctx, 9B model)
f16 (default)	~16 GB
q8_0	~8 GB
q4_0	~4 GB

On an 8 GB Mac, use q4_0 KV cache and reduce context to -c 32768 (32K). On 16 GB, you can comfortably do 128K context. On 32 GB+, you can run larger models or multiple parallel slots.

If you’re still running out of memory, reduce context size first (-c), then try a smaller quantization (Q3_K_M instead of Q4_K_M).

Test it

curl -s http://localhost:8080/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "Qwen3.5-9B-Q4_K_M.gguf",
    "messages": [{"role": "user", "content": "Hello!"}],
    "max_tokens": 50
  }' | jq .choices[0].message.content

Get the model name

If you forget the model name, query the models endpoint:

curl -s http://localhost:8080/v1/models | jq '.data[].id'

---

Option B: MLX via omlx

omlx is a macOS-native app that manages and serves MLX models. MLX is Apple’s own machine learning framework, optimized specifically for Apple Silicon’s unified memory architecture.

Install

Download and install from omlx.ai. It provides a GUI for model management and a built-in server.

Download the model

Use the omlx app to browse and download models. Search for Qwen3.5-9B-mlx-lm-mxfp4 and download it. Models are stored locally (typically in ~/.omlx/models/).

Start the server

omlx serves models on http://127.0.0.1:8000 by default. Start serving from the app UI, or use the CLI if available.

Test it

curl -s http://127.0.0.1:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "Qwen3.5-9B-mlx-lm-mxfp4",
    "messages": [{"role": "user", "content": "Hello!"}],
    "max_tokens": 50
  }' | jq .choices[0].message.content

List available models

omlx can serve multiple models simultaneously:

curl -s http://127.0.0.1:8000/v1/models | jq '.data[].id'

---

Benchmarks: llama.cpp vs MLX

Both backends tested on the same machine (Apple M5 Max, 128 GB unified memory) running the same model (Qwen3.5-9B) at comparable quantization levels (Q4_K_M for GGUF, mxfp4 for MLX). Five diverse prompts, three runs each, backends tested sequentially to avoid resource contention.

Results

Metric	llama.cpp (Q4_K_M)	MLX (mxfp4)	Winner
TTFT (avg)	67 ms	289 ms	llama.cpp (4.3x faster)
TTFT (p50)	66 ms	286 ms	llama.cpp (4.3x faster)
Generation (avg)	70 tok/s	96 tok/s	MLX (37% faster)
Generation (p50)	70 tok/s	96 tok/s	MLX (37% faster)
Total time (512 tokens)	7.3s	5.5s	MLX (25% faster)

What this means

• llama.cpp excels at prompt processing — its flash attention + quantized KV cache pipeline gets you the first token in ~66ms. If you’re building interactive applications where perceived responsiveness matters (chatbots, autocomplete), this is a meaningful advantage.

• MLX generates tokens ~37% faster once it gets going. For batch workloads, long-form generation, or any task where total completion time matters more than initial latency, MLX finishes sooner.

• Both backends are extremely consistent — variance across runs was negligible. You can rely on these numbers.

Which one should you pick?

Use case	Recommendation
Interactive chat, low-latency tools	llama.cpp
Long-form generation, bulk processing	MLX (omlx)
Memory-constrained (8-16 GB)	llama.cpp (quantized KV cache is unmatched)
Serving multiple models simultaneously	omlx (built-in multi-model support)
Maximum compatibility (Linux too)	llama.cpp

---

Connect to Hermes

Once your local server is running:

hermes model

Select Custom endpoint and follow the prompts. It will ask for the base URL and model name — use the values from whichever backend you set up above.

---

Timeouts

Hermes automatically detects local endpoints (localhost, LAN IPs) and relaxes its streaming timeouts. No configuration needed for most setups.

If you still hit timeout errors (e.g. very large contexts on slow hardware), you can override the streaming read timeout:

# In your .env — raise from the 120s default to 30 minutes
HERMES_STREAM_READ_TIMEOUT=1800

Timeout	Default	Local auto-adjustment	Env var override
Stream read (socket-level)	120s	Raised to 1800s	HERMES_STREAM_READ_TIMEOUT
Stale stream detection	180s	Disabled entirely	HERMES_STREAM_STALE_TIMEOUT
API call (non-streaming)	1800s	No change needed	HERMES_API_TIMEOUT

The stream read timeout is the one most likely to cause issues — it’s the socket-level deadline for receiving the next chunk of data. During prefill on large contexts, local models may produce no output for minutes while processing the prompt. The auto-detection handles this transparently.