Magnitude 0.0.13: Pilot Hardware-Aware Local Models, Keep a Manual Rollback

OpenSourceChoice verdict: Magnitude is worth a bounded developer pilot if model selection and agent configuration consume more time than local inference itself. Do not make it the only path to critical coding work yet. Pin version 0.0.13, keep Ollama or direct llama.cpp as a rollback, and treat the loopback service as a security boundary—not as proof that local webpages are safely isolated.

This researched analysis was checked on September 8, 2026. We reviewed current documentation, repository and package activity, releases, licensing, issues and independent interest signals. We did not install Magnitude, run its models or reproduce the reported defects.

The decision at a glance

Question	Assessment
Best fit	Developers with supported Mac or Linux hardware who already use an agent and want guided model selection
Poor fit	Production services, native Windows-only teams, or environments requiring a documented security-response process
Current stable package	`@magnitudedev/cli` 0.0.13, published September 8
Main value	Hardware profiling, ranked model configurations and automatic harness connections
Main risk	A young `0.0.x` control layer with unresolved security, compatibility and performance reports
Real cost	Hardware, model storage, electricity, cold-load delay, upgrades and operator time
Parallel alternative	Ollama for simpler distribution; llama.cpp for direct control and a performance baseline

Why Magnitude matters now

Magnitude reached GitHub's weekly Trending list with 2,430 stars added during the observed week and 4,064 total. The npm downloads API counted 3,612 CLI downloads from August 31 through September 6. These numbers establish current attention, not reliability.

Version 0.0.13 is also a meaningful trigger. Its only release-note change adjusts how the macOS background service runs to prevent throttling. That may address a report in which the bundled engine was much slower than llama.cpp on the same M4 and GGUF file, but the issue was still open awaiting a retest when checked. The correct conclusion is “measure the fix,” not “performance is fixed.”

What it does—and what it does not

Magnitude profiles processor, memory, architecture and acceleration, then ranks model, quantization and context combinations. It downloads a selected model, runs a Rust inference service built on llama.cpp, and connects supported harnesses including Codex, Claude Code, OpenCode, Cline and Pi.

The service listens on 127.0.0.1:10100 and exposes OpenAI-compatible and Anthropic-compatible APIs. Models load on demand, can unload when idle or under memory pressure, and remain stored under ~/.magnitude. This is more opinionated than the runner layer in our local GPT stack guide: Magnitude tries to choose and configure the complete agent-facing path.

It does not make a small local model equal to a frontier hosted model. It does not remove model-license review, prove that estimated tokens per second match a workload, or provide a documented container deployment. Windows support currently means WSL, while the official release assets target macOS and Linux.

Privacy is local, but the boundary needs work

Ordinary inference can keep prompts and files on the machine. After packages and model weights are downloaded, the documented inference path can run offline. The optional Magnitude Harness web search uses an EXA_API_KEY, and installation, model acquisition and updates still require network access. Session data, logs, models and configuration persist locally, so disk encryption, backup scope and deletion rules still matter.

The most important current security signal is open issue 78. A static source review reported that the local server accepts CORS origins null and file://, not only localhost origins. A maintainer agreed those entries should be removed and suggested a dedicated application origin. We did not reproduce the behavior, but a loopback listener reachable from an untrusted page is enough reason to keep sensitive repositories out of the first pilot and monitor the issue.

The repository had no SECURITY.md and no published repository advisories when checked. That is not evidence of safety or compromise; it means adopters lack a documented private reporting and disclosure contract. Fine-grained local-agent permissions are also an open proposal. Local model privacy and safe shell authority are separate decisions.

Maturity and operational cost

The Apache-2.0 license is straightforward for the Magnitude code, but downloaded model weights keep their own licenses. The project is actively maintained: 0.0.11, 0.0.12 and 0.0.13 arrived within six days, and commits continued on the publication date.

Open reports show the current edge cases: a strict Mistral chat template can break an agent session after a tool call; a non-Ubuntu installation remains unresolved; and the M4 performance report needs confirmation on 0.0.13. “No token bill” still means storage, RAM or VRAM, download bandwidth, electricity, cold starts and time spent qualifying every model/harness combination.

Teams that only need a familiar local endpoint should compare Ollama first. Teams that need knobs, portability or an independent speed baseline should compare direct llama.cpp. Apple-only operators needing a multi-client server can also use our oMLX adoption analysis. If the real requirement is the agent rather than the runner, shortlist from the open-source Codex alternatives before changing inference infrastructure.

A measurable 10-day pilot

Use one non-sensitive repository and one reversible task set. Keep the existing runner installed and record its configuration before Magnitude changes any harness connection.

Gate	Proceed only if
Quality	The local model completes at least 80% of a labeled 20-task set without unsafe or invented edits
Performance	p95 first-token latency and completion time stay within the team's declared interactive limit
Compatibility	Chat, tool calls, cancellation and long contexts pass in every selected harness/model pair
Security	The service remains loopback-only; untrusted origins and another local account cannot use it
Privacy	Network observation finds only approved package, model and optional search destinations
Reliability	Load, unload, restart, low-memory recovery and a full workday soak produce no stuck sessions
Reproducibility	A second machine can install the pinned CLI and exact model revision from the runbook
Exit	Removing the connection restores the previous runner without losing repository or session data

Reject the rollout if the recommendation estimate is the only performance evidence, if tool calls fail on the chosen model, or if the team cannot explain which process can reach port 10100. Success means Magnitude reduces setup and switching work while the same tasks remain correct, private and reversible.

Conclusion

Magnitude addresses a real gap between “install a local model runner” and “make several coding agents use the right model for this machine.” Its hardware-aware catalog, on-demand loading and harness connections are useful enough to test now.

Its current versioning and open issues argue against trust by convenience. Pin 0.0.13, benchmark against llama.cpp or Ollama, keep the first repository non-sensitive, and promote only after the loopback, tool-call and recovery gates pass. The product earns a pilot; the evidence does not yet earn a default platform.

Sources and research record

Sources were accessed September 8, 2026. Dynamic counts are observations from that date. Project benchmarks and issue reports were not treated as independent hands-on verification.