Magic, the coding-focused AI lab, published research claiming its pretraining recipe now needs roughly a tenth of the compute that leading open-weight base models require to reach comparable quality. The company says it matched DeepSeek V4 Pro’s base model using about 50 times fewer training FLOPs, then kept scaling and surpassed several open base models outright. That claim matters because Magic has never shipped a public model, so the only evidence for its approach is the comparison it chose to publish.
The metric behind the headline is bits-per-byte loss, a way of scoring next-token prediction on held-out text that sidesteps the prompt sensitivity of benchmark-style evals on raw base models. Magic ran this test against the latest open-weight base checkpoints it could obtain from DeepSeek, Moonshot’s Kimi line, and Nvidia’s Nemotron 3 Ultra, using a 6ND approximation (parameter count times token count) rather than exact training FLOPs, since exact figures aren’t published for most of those models. Base models from Anthropic, Google, and OpenAI are never released, so none of them appear in the comparison at all.
Magic’s own numbers show the gap is not settled even against open models. The post notes that Moonshot’s newer Kimi K3 and Meta’s Muse Spark reportedly beat Kimi K2, the checkpoint Magic benchmarked against, by 2.5x and 3.3x on similar measures. If those newer releases hold up, part of the multiple Magic is claiming may already be closing before the ink dries, since a comparison against last quarter’s open models says less about the current frontier than it does about a moving target.
On process, Magic describes real verification work: it ran its held-out logprob evaluations across both vLLM and SGLang inference backends, found inconsistencies in some backends, and brought in Fireworks to cross-check baseline numbers independently. That is more rigor than a typical lab blog post offers. It does not, however, change the fact that the underlying training runs, the data mixture, and the eval sets were all built and scored by Magic itself, with no outside party able to rerun the comparison on the same infrastructure.
On substance, Magic says the recipe change was not one breakthrough but the compounding effect of dozens of adjustments to model architecture, optimizer choice, training objective, and data curation, validated by training small models across two orders of magnitude of compute and extrapolating with power-law fits before committing to a full-scale run. It also ran a short reinforcement-learning pass, starting directly from the base model with no supervised fine-tuning, and reported strong pass rates on a private, held-out competition math set. The company frames long context and pretraining as two of three pillars, alongside agentic RL, that it believes are sufficient to build coding agents that outperform humans and to automate AI research itself.
That strategic bet deserves to be taken on its own terms, separate from whether the 10x figure holds exactly. A lab without a frontier-scale compute budget genuinely has only one lever left: get more capability out of every FLOP it can afford. Efficiency claims from a company that is not selling API access to the model in question are structurally hard to audit. Nobody outside Magic can rerun the eval suite, and the claim does double duty as a fundraising and recruiting pitch whether or not the multiple survives scrutiny from a rival lab’s own numbers.
For operators, the practical read is narrower than the headline: treat the 10x figure as Magic’s internal accounting until an independent party benchmarks a released Magic model, and watch whether the company’s next update addresses the Kimi K3 and Muse Spark comparisons directly.
Magic detailed the pretraining recipe and benchmarks in a research post on its own blog, magic.dev, published this month.
