Mirendil says one researcher pretrained a coding model for under $10 million

What happened

On October 21, 2026, Mirendil published Mirendil-Base. The company says one researcher built it in 3.5 months, with less than $10 million in compute. That figure is all-in: the research, the small experiments, the final run, the agents' token spend, and data processing. Mirendil says there was no distillation and no purchased or licensed data. Agents wrote over 99 percent of the code, ran the experiments, and diagnosed failures. The researcher decided what to try. The code was written from scratch on PyTorch and JAX. Training data was public code, web pages, and public datasets. The post puts Mirendil-Base in the 100 to 500 billion parameter range. A larger model, 500 billion to 2 trillion, is about 80 percent through training. A third, 2 to 10 trillion, is about 19 percent through. On seven external code repositories, Mirendil's bits-per-byte average is 0.222. The same table lists Kimi-K3 at 0.193, Nemotron-550B-Base at 0.247, GLM-5.3 at 0.242, and DeepSeek-V4-Pro-Base at 0.243. Lower is better.

Why it matters

The claim is not that the model leads. On Mirendil's table it does not: Kimi-K3 is ahead, and Mirendil notes that Kimi-K3 and GLM-5.3 are post-trained, which usually looks slightly worse on this metric than a base model. The claim is that the team was one person. If that account holds, the scarce input for a coding base model of this size is compute and judgment, not a pretraining department. The page does not include a weight file or an API. Post-training has started. Two larger runs are still going.

Mirendil lists the comparison models as mixture-of-experts systems: Kimi-K3 at 2.8 trillion total and 104 billion active, Nemotron-550B-Base at 550 billion and 55 billion, GLM-5.3 at 753 billion and 40 billion, DeepSeek-V4-Pro-Base at 1.6 trillion and 49 billion, and Inkling at 975 billion and 41 billion. Inkling is on the first figure and not in the seven-repository table.

The first figure is held-out code from Mirendil's own repositories. The seven-repository table is code published after the models were trained. Mirendil says a post-trained model generally scores slightly worse on bits per byte than its base version, so Kimi-K3 and GLM-5.3 are not a like-for-like base comparison.

Mirendil says the next model will be built partly by Mirendil models, and that this is not fully autonomous recursive self-improvement. That sentence is the company's, not a demonstration that the loop has closed.