Ollama vs llama.cpp

full-guide.md

these two get pitched as rivals, but they're more like a car and its engine. llama.cpp is the c/c++ inference engine: it runs open GGUF models on your own hardware, pulls them straight from Hugging Face and serves them over an API. Ollama wraps local models in one command and a background service with a REST API on localhost:11434. LM Studio puts a desktop app on top and runs models via llama.cpp and MLX.

on our local models board llama.cpp is #1 (9.0/10), Ollama #2 (8.5) and LM Studio #3 (7.5). all three are free for local use.

speed: llama.cpp, by a nose

two independent tests (Mozilla.ai in september 2026, InventiveHQ in june 2026) had llama.cpp as the fastest of the three. InventiveHQ clocked Ollama 10 to 14% slower than raw llama.cpp. LM Studio was level with llama.cpp on an NVIDIA card but 29% behind on a Mac.

the honest bit: on a laptop that's the difference between quick and slightly less quick. if Ollama gets you running in two minutes and llama.cpp costs you an afternoon of build flags, the afternoon is the bigger cost. speed matters when you're serving a model to an app or an agent all day.

the setting that quietly breaks coding agents

Ollama picks a default context length from your VRAM: 4k tokens under 24 GiB, 32k between 24 and 48 GiB, 256k at 48 GiB and up. its own docs say agents and coding tools want at least 64,000 tokens. so on a normal laptop GPU, point a coding agent at Ollama without changing anything and long prompts get cut short. fix it with the slider in the app's settings, or OLLAMA_CONTEXT_LENGTH=64000 ollama serve, then check the CONTEXT column in ollama ps.

llama.cpp's server defaults to the context the model was trained with (-c 0), so it doesn't bite in the same way. more memory, mind.

APIs: both talk OpenAI

• Ollama: OpenAI-compatible endpoints under /v1: chat completions, the responses API (stateless only), embeddings and model listing, plus its own REST API.
• llama.cpp: llama serve gives you OpenAI-compatible chat completions, responses and embeddings, and it also speaks the Anthropic messages API.
• LM Studio: a local server with OpenAI- and Anthropic-style APIs.

so for most apps it's a base URL change, whichever you run.

hardware and control: llama.cpp

this is why it's #1. Apple silicon is first-class (metal), plus CUDA for NVIDIA, HIP for AMD, vulkan, SYCL and a long list of others. it'll split a model across CPU and GPU when it's bigger than your VRAM, and it supports quantisation from 1.5-bit to 8-bit. MIT licensed, no account, no paid tier. the price is your time: you're choosing build options and flags yourself.

convenience and cloud: Ollama

one command to pull and run a model, and running models on your own hardware is always unlimited. the extra trick: optional cloud models through the same tool when your laptop runs out of puff. local stays free. the plans, for the record: Free $0 (local use unlimited); Pro $20/mo ($200/yr) with $60 of cloud credits a month; Max $100/mo with $300 of credits; Team $500/mo (early access). the paid ones only buy cloud usage. Ollama says cloud prompts and responses are never logged or trained on.

where LM Studio fits

it's the one with a proper GUI: search for models, download them, chat with them and serve them. good for trying models side by side before you wire one into an app. the catch is that the app itself isn't open source, and it was the slowest of the three on a Mac in the tests above.

plugging any of them into a coding agent

OpenCode has ready-made configs for all three in its docs, which is the easiest way to run a coding agent on local models. Claude Code won't, it's Claude models only: see OpenCode vs Claude Code. whichever runner you use, check the context length first.

and which model?

we rank the runners, not the weights. the best local model for coding depends on how much VRAM you've got, and it changes every few weeks, so we're not going to pretend otherwise with a made-up table.

who should pick which

• want it working in two minutes, behind an API: Ollama
• want the fastest inference and full control: llama.cpp
• want a GUI to browse and compare models: LM Studio
• on a Mac and chasing speed: llama.cpp
• want cloud models from the same CLI when local isn't enough: Ollama
• want to run a coding agent locally: any of the three behind OpenCode, with context set to 64k or more

faq

is Ollama faster than llama.cpp?

no. in two independent tests llama.cpp was the fastest of the three runners, and Ollama ran 10 to 14% slower than raw llama.cpp in one of them.

is llama.cpp free?

yes. it's MIT licensed, with no account and no paid tier. you run it on your own hardware.

is Ollama free?

yes for local use, with no limits. paid plans only add cloud model usage.

Ollama vs LM Studio: which is better?

on our board Ollama ranks above LM Studio. Ollama is the better fit for wiring a model into an app or agent. LM Studio is the better fit if you want a desktop app to find and test models.

why does my local model get dumb in a coding agent?

often it's the context length. Ollama defaults to 4k tokens on GPUs under 24 GiB of VRAM, and its docs recommend at least 64,000 for coding tools.

sources

checked 2026-10-20

• llama.cpp README (quick start, backends, quantisation, licence): https://github.com/ggml-org/llama.cpp
• llama.cpp server docs (OpenAI and Anthropic compatible routes, context default): https://github.com/ggml-org/llama.cpp/tree/master/tools/server
• Ollama pricing (local use unlimited, cloud plans, data): https://ollama.com/pricing
• Ollama context length docs: https://docs.ollama.com/context-length
• Ollama OpenAI compatibility docs: https://docs.ollama.com/api/openai-compatibility
• Ollama licence (MIT): https://github.com/ollama/ollama
• LM Studio pricing: https://lmstudio.ai/pricing
• OpenCode providers docs (Ollama, LM Studio and llama.cpp configs): https://opencode.ai/docs/providers/
• speed tests as recorded in our data: https://blog.mozilla.ai/benchmarking-local-llm-servers-llama-cpp-llamafile-lm-studio-and-ollama/ and https://inventivehq.com/blog/ollama-vs-lm-studio-vs-llama-cpp
• ranks, scores, verdicts and plans: numptr data/tools.json (local models board)

more guides