What is Magnitude?
Magnitude runs open models on your own computer and connects them to the coding agent you already use. When you download a model it spends up to a minute on your chip trying different ways to split the work across GPU threads, and keeps the fastest combination that does not change the output. Each supported model ships with a speculative-decoding drafter (MTP, DFlash, DSpark or LiLiCorr), and sessions share prefix caches.
The engine is written in Rust on Magnitude's own compiler and runtime, Seismic, which targets CPUs, Apple Metal, NVIDIA CUDA (Ampere and newer) and Vulkan; there is no ROCm backend. It ships as a desktop app and CLI for macOS, Windows and Linux, serves OpenAI- and Anthropic-compatible APIs on port 10100, and configures Pi, OpenCode, Hermes, OpenClaw, Codex, Claude Code, Oh My Pi and Cline in one click.
The repository was created on 12 June 2026 by Magnitude AI Inc., is Apache-2.0 licensed, and released CLI 0.2.13 on 28 October 2026, its third release in three days.
Is Magnitude worth it?
Magnitude can be faster than the engines it targets, but how much depends heavily on the machine, and every headline number is the vendor's own. At its Launch HN for Y Combinator's S25 batch in early October 2026, the founders claimed up to twice llama.cpp's speed, based on Qwen 3.6 35B A3B at 4-bit with a 64K-token prompt: decode rose from 30 to 57 tokens a second on an M4 Pro Mac (92%), but only from 49 to 58 on a DGX Spark (19%), and prefill gained 9% and 23%. The comparison was not like for like: the founders said they ran llama.cpp with a 16-bit KV cache because its compressed cache slowed decode, while Magnitude used its own compressed cache.
The benchmark page now compares against Ollama v0.35.1 instead, across 11 models on a 48 GB M4 Pro Mac mini and a DGX Spark: average decode 2.19 and 2.17 times faster, prefill about 1.3 times. The spread is wide, from 4.84 times faster (Qwen 3.5 4B decode on the Spark) to 0.76 times, slower, on Nemotron 3.5, and the page warns that Ollama's weak drafting on three Gemma models inflates those gains and that answer quality was not scored. Users on the Hacker News thread posted the other side: on an M5 Max one found llama.cpp about twice as fast as Magnitude 0.2.1, another measured rapid-mlx at 175 tokens a second against Magnitude's 161, and on an RTX 5070 Ti llama.cpp decoded 20 to 30% faster. The founders said M5 Macs are a known weak spot because the kernels do not yet use Metal 4's matrix hardware.
The local defaults are sensible. The API listens on 127.0.0.1:10100 with no key for the same machine; turning on network access requires a generated key, slows guessing after five wrong attempts and still speaks plain HTTP, so the docs recommend Tailscale, SSH or an HTTPS proxy. In our reading of the code, its OpenTelemetry traces go to a collector on 127.0.0.1. The Claude Code connection rewrites ~/.claude/settings.json to send every request, including hosted Claude ones, through Magnitude's local gateway, which you must keep running and later undo.
Where it falls short of Ollama and llama.cpp is reach. CUDA builds skip pre-Ampere cards, AMD and Intel GPUs go through Vulkan only, and speed depends on hand-written kernels for the popular model families Magnitude chooses to support. Two developers wrote almost all of the code, and 30 issues are open, including installs stuck at model assessment and failures on non-Ubuntu Linux.
Where it earns its place
A developer on a 48 GB Apple silicon Mac using OpenCode
Get quicker local completions than Ollama gives on the same model
Install the app, take Discover's Balanced pick and click Connect for OpenCode. On the vendor's M4 Pro test decode averaged 2.19 times Ollama's speed; check your own model in Usage, since three Gemma results on the Mac were inflated by Ollama's weak drafting.
A three-person studio with one RTX 4090 workstation
Share one local model with every laptop in the office
Run `magnitude server setup` so it starts at boot and updates itself, turn on network access and give each laptop the API key. The link is plain HTTP, so put it on Tailscale or behind an HTTPS proxy; only inference is available remotely.
An engineer on a Claude Code team who wants an offline fallback
Run Claude Code against a local model when the plan's limits run out
Connect Claude Code from Magnitude and start it with `claude --model anthropic-local/<id>`. The connection rewrites ~/.claude/settings.json and routes hosted requests through the local gateway too, so disconnect and restart Claude Code before using it on its own again.
Before you commit
- Setup
- `curl -fsSL https://magnitude.dev/install.sh | sh`, the PowerShell equivalent, or .dmg, .exe, .deb, .rpm and pacman installers. Each model download adds up to a minute of kernel tuning before first use.
- Integrations
- OpenAI-compatible API at 127.0.0.1:10100/inference/v1 and Anthropic-compatible at /inference/anthropic; one-click setup for Pi, OpenCode, Hermes, OpenClaw, Codex, Claude Code, Oh My Pi and Cline. Backends are CPU, Metal, CUDA (Ampere and newer) and Vulkan.
- Data and privacy
- Prompts and models stay local according to the docs; models live in ~/.magnitude/models, and OpenTelemetry traces go to a collector on 127.0.0.1 in our reading of the code. No key is needed on the same machine; network access requires one and is plain HTTP. Apache-2.0, with each model under its own licence.
- Cost at scale
- No paid edition is documented, so hardware is the cost. The vendor's benchmarks used a 48 GB Mac mini with an M4 Pro and a 128 GB DGX Spark; the server mode suits one shared workstation rather than high-concurrency serving.
- Limits
- 30 open issues, including stuck model assessment (#142) and installs on non-Ubuntu Linux (#22); three CLI releases in three days; no ROCm and no pre-Ampere CUDA; speed depends on kernels written for the model families it supports.
What's good
- Per-machine kernel tuning plus a drafter for every supported model
- Open-source benchmark harness and per-model results
- Localhost by default; network access needs a key
- One-click setup for Claude Code, Codex, OpenCode, Cline and others
What's not
- Speed claims are self-reported; users found llama.cpp or MLX faster on M5 Max and RTX 5070 Ti
- Slower than Ollama on some models, such as Nemotron 3.5 decode on DGX Spark
- No ROCm; CUDA builds need an Ampere or newer card
- Claude Code integration routes all requests through its gateway
Best for
Developers on Apple silicon or recent NVIDIA cards who want a fast local model behind a coding agent with little setup.
Magnitude alternatives
Magnitude: common questions
- Is Magnitude worth it?
- A smart idea with real speed behind it on some machines, but the 2x headline is the company's own measurement and has not held up on every user's hardware. We rate it 3.6 out of 5.
- How much does Magnitude cost?
- Free and open source under Apache-2.0. No paid edition is documented; model licences are separate.
- What are the downsides of Magnitude?
- Speed claims are self-reported; users found llama.cpp or MLX faster on M5 Max and RTX 5070 Ti; slower than Ollama on some models, such as Nemotron 3.5 decode on DGX Spark; no ROCm; CUDA builds need an Ampere or newer card; claude Code integration routes all requests through its gateway.
- Who is Magnitude best for?
- Developers on Apple silicon or recent NVIDIA cards who want a fast local model behind a coding agent with little setup.
- How much does Magnitude cost for a team or in production?
- No paid edition is documented, so hardware is the cost. The vendor's benchmarks used a 48 GB Mac mini with an M4 Pro and a 128 GB DGX Spark; the server mode suits one shared workstation rather than high-concurrency serving.
- Is Magnitude safe for company data?
- Prompts and models stay local according to the docs; models live in ~/.magnitude/models, and OpenTelemetry traces go to a collector on 127.0.0.1 in our reading of the code. No key is needed on the same machine; network access requires one and is plain HTTP. Apache-2.0, with each model under its own licence.
- What are the best Magnitude alternatives?
- The closest alternatives we have reviewed are Ollama, llama.cpp, MLX and Jan.
Answers checked Oct 28, 2026.