# What Exactly Did Magnitude Launch, and Who Built It?

Magnitude is a YC S25 company founded by Anders and Tom, both software engineers who previously built an open source browser agent that reached 4k+ GitHub stars and 100k+ downloads, according to the founders' Hacker News launch post. Their stated motivation is specific: they wanted to run their browser agent on local models and found no inference engine worked for their use case. The product is a self-optimizing inference engine that runs on Mac, Linux, and Windows across any hardware. The company claims it is up to 2x faster than llama.cpp. The launch post does not name the specific models tested, the context lengths used, the hardware configurations, or whether the comparison was against llama.cpp's default build or a tuned one. That omission matters. A 2x claim against llama.cpp is a strong claim against a mature, heavily optimized codebase with years of community tuning. The founders' prior work gives them credibility on agent workloads, but credibility is not a benchmark.

# Why Does an Agent-Specific Inference Engine Matter?

According to the Magnitude founders, the trigger for building the engine was that no existing inference engine worked for their browser agent use case. That is a design-target statement, not a performance statement, and it is the more interesting half of the launch. Agent workloads differ from chat workloads in ways that stress inference engines differently. An agent makes repeated tool calls, re-sends growing context, and often runs many short generations rather than one long one. KV cache management, prompt reuse, and scheduling across turns matter more than raw tokens-per-second on a single prompt. If Magnitude's self-optimization is real, it is likely optimizing for that loop shape rather than beating llama.cpp at a standard single-prompt benchmark. The founders did not publish a breakdown of where the 2x comes from, so this remains inference on my part, not a sourced claim.

# How Credible Is the 2x Claim Without a Public Benchmark?

The Hacker News launch post states the engine is "up to 2x faster than llama.cpp." The phrase "up to" is doing real work here. It permits a best-case configuration that may not represent typical use. Magnitude reported that the engine works on Mac, Linux, and Windows on any hardware, which is a broad compatibility claim that, if true, is genuinely difficult to achieve. Cross-platform inference with hardware-specific optimization usually forces tradeoffs, and the launch post does not explain how those tradeoffs were handled. The GitHub repository is the primary source, and as of the launch post, the public evidence for the 2x figure is the founders' own statement. That is not disqualifying, but it is not verification either. Independent reproduction is the only thing that settles this.

# How Does Magnitude Compare to the Incumbents?

Dimension	Magnitude	llama.cpp	Ollama
Primary design target	Agent loops (multi-turn, tool calls)	General local inference	Developer-friendly local serving
Claimed performance	Up to 2x vs llama.cpp (unverified)	Baseline reference	Wraps llama.cpp
Platform support	Mac, Linux, Windows	Mac, Linux, Windows	Mac, Linux, Windows
Self-optimization	Core claimed feature	Manual flags and tuning	Limited auto-config
Maturity	Launch-stage, YC S25	Years of community hardening	Widely adopted wrapper
Verdict	Watch closely for agent workloads; unproven vs llama.cpp on general inference	Still the default until independent benchmarks say otherwise	Not the comparison that matters here

# What Does This Mean for Local Agent Developers?

The practical question for a developer running an agent locally is not which engine wins a synthetic benchmark. It is which engine keeps a multi-turn agent loop responsive on a laptop without thermal throttling. Magnitude's founders built a browser agent before building an inference engine, which means they are optimizing for a workload they have actually shipped. That is a stronger signal than a generic speed claim, because it suggests the optimization targets are drawn from real traces rather than synthetic prompts. Ollama, by contrast, is a serving wrapper around llama.cpp rather than a competing engine. If Magnitude's claims hold, the competitive pressure lands on llama.cpp directly, not on Ollama, which can adopt whatever wins.

# Predictions

1. Magnitude will publish a public, reproducible benchmark with named hardware and models within 60 days of its Hacker News launch, or the 2x claim will stop being cited in developer discussions by Q1 2027. 2. llama.cpp maintainers will ship agent-specific tuning, likely KV cache reuse improvements for multi-turn loops, within two release cycles, in direct response to launches like Magnitude. 3. At least one major local-agent framework will add Magnitude as a supported backend within six months, treating it as a drop-in alternative to llama.cpp rather than a replacement.

# Article Summary

• Magnitude's real differentiator is agent-loop specialization, not the unverified 2x speed claim.
• The 2x figure is "up to" language with no published methodology, models, hardware, or context lengths.
• Ollama is not the competitor here; it wraps llama.cpp and can adopt any winning engine.
• The founders' prior browser agent work gives them credible insight into the workload they are optimizing for.
• The claim's survival depends on reproducible benchmarks within roughly 60 days.