Magic Matches DeepSeek V4 Pro Base Using 50x Less Compute

Frontier pretraining has become a game reserved for labs with 100,000-chip clusters. Magic, a startup building coding and autonomous R&D models, just posted numbers suggesting algorithmic efficiency can substitute for hardware. Its new recipe matches DeepSeek V4 Pro Base while burning roughly 50x less compute, then scales up 10x further to beat every publicly available open base model.

The headline numbers

Magic reports matching DeepSeek V4 Pro Base using roughly 50x fewer FLOPs, about half the pretraining compute used for GPT-3, or approximately $0.5M on GB200 hardware. Scaling 10x further (around $4M) meaningfully outperformed all publicly available open base models on perplexity evals. Under DeepSeek V4 Pro's own recipe, the scaling laws in Magic's report imply that reaching the same capability level would cost more than $100M, before considering whether enough training data even exists to get there.

How they measured it

Base models are notoriously slippery to benchmark: without RL or SFT, they are extremely prompt-sensitive. Magic sidesteps this by measuring bits-per-byte loss on held-out data, a metric that normalizes across tokenizers and requires no prompt engineering. They then fit a Chinchilla-style scaling law to project the compute needed to reach any given capability level.

To stress generalization rather than memorization, their eval sets were built from sources unlikely to appear in training data:

• Code evals drawn from Magic's own codebase and private codebases acquired from other startups.
• Reasoning evals built by generating chain-of-thought walkthroughs for private math problems using Kimi K3, filtered for correct answers.
• Recent, low-citation research papers for text and knowledge evals.
• Aggressive decontamination: vendored OSS code and any document with a matching 96-character window or high Jaccard similarity to training data was removed.

Logprob measurements were validated across vLLM and SGLang on GB200 and GB300. The team found bugs in some backends along the way and partnered with Fireworks to double-check baselines in a third inference engine. The eval parser was deliberately different from the pretraining parser, since models can otherwise learn the quirks of their own preprocessor.

Dozens of small changes, compounded

No single optimizer trick or exotic architecture drives these results. Progress came from tens of changes compounded across model architecture, optimizer, training objective, and data curation. The team's starting point was a stable foundation: smooth convergence, low-precision training quality equivalent to FP32, clean infrastructure, correct hyperparameter scaling rules, and aggressive bug-hunting.

Two methodological details stand out for anyone running their own scaling research:

1. NanoGPT speedruns give fast feedback, but many improvements to tiny models do not transfer to large ones, and some features present in most LLMs can be removed without harming large-scale performance.
2. For each change, they train three models spanning two orders of magnitude of compute and keep the change only if its power-law fit predicts it will help at scale. Every few weeks they push to 1/10th of the hero scale; every few months they run a full hero run.

Domain-level data tracking

Because Magic targets software engineering and autonomous AI R&D, they explicitly rebalance data mixes and track what each rebalancing costs. Their held-out research text eval breaks down by subject, with granular buckets covering specific tool documentation and key alignment papers, giving precise per-domain loss signals.

To prevent the model from memorizing eval documents that legitimately belong in the training set, Magic rewrites and summarizes those documents using a third-party frontier LLM and creates fresh knowledge evals once per model generation. Domains deprioritized on purpose include local news, sports events, and facts about public figures.

Does the pretrain quality survive RL?

Base-model perplexity only matters if it carries over after post-training. Magic ran a short math RL run with a 16k chain-of-thought budget, starting directly from the base model with no SFT or distillation. On a private competition math eval harder than AIME, their V5 (e24) model hit 72% pass@1, ahead of DeepSeek V4 Pro at 65% and behind Kimi K3 at roughly 75%, while using orders of magnitude less RL compute than either baseline.

The e24 model's AIME26 pass@1 first crossed 90% (and 100% pass@16) at about 0.2% of its pretraining compute budget, using just over 6k chain-of-thought tokens on average.

Where this leaves the ecosystem

The prevailing assumption for the past two years has been that only labs with nine-figure compute budgets can reach the frontier. Magic's data suggests the efficiency frontier is moving faster than the hardware frontier. For reference, Kimi K3 and Meta's Muse Spark reportedly show 2.5x and 3.3x gains over Kimi K2; Magic's claimed multiplier over the same baseline is an order of magnitude larger.

A footnote observation worth noting: Nemotron 3 outperforms DeepSeek V4 Pro on base-model bits-per-byte across domains and inference engines, which suggests its weaker post-RL benchmark results may reflect a post-training problem rather than anything wrong with the pretrain.

What comes next

Magic considers pretraining and long-context work mature internally. The next phase focuses on:

• Scaling long-horizon RL so agents keep learning after deployment through long context.
• Exploration and credit assignment in long-horizon RL, plus the systems infrastructure to scale it.
• Alignment training against narrowly elicited latent knowledge, targeting stronger theoretical guarantees.
• Further pretraining improvements, and eventually a public model release.

Magic also claims to be likely the smallest team in the world training trillion-parameter models. If the numbers hold under external scrutiny when a model ships, the assumption that frontier pretraining requires a 100k-chip cluster and a thousand researchers will need serious revision.