Mirendil-Base: one researcher, agents, <$10M
Partly confirmed
The takeaway
On Oct 21, 2026 Mirendil, the ex-Anthropic startup working on self-improving AI, announced Mirendil-Base, a 100B–500B-parameter coding base model.
Status
- Claim
Partly confirmed
- Our reporting
- Medium confidence
- Importance
- 3 of 5
- Last verified
- 26 October 2026
Your AI and this story
- GPT-6.2 Astra143 days after its cutoff
- Claude Opus 5.5113 days after its cutoff
- Gemini 4 Hydrogen204 days after its cutoff
- Grok 4.7143 days after its cutoff
None of these four assistants can know about it. The closest, Claude Opus 5.5, stops 113 days before it.
Key facts
- Family: Mirendil-Base (100B–500B parameters), Base-Large (500B–2T, ~80% through training) and Base-XL (2T–10T, ~19%), described as ‘coding-specialized models’. Post-training has started
- Cost and team: under $10M all-in (agent tokens, data processing, research, small experiments and the final run) over 3.5 months, with one unnamed researcher. Agents wrote over 99% of the code; the researcher chose what to try and what to trust
- Data: only public data (GitHub and other public code, web pages, public datasets); no purchased or licensed data, no distillation and no synthetic data. Code written from scratch on PyTorch and JAX
- Bits per byte on seven external repositories (lower is better): Kimi-K3 (post-trained) 0.193, Mirendil-Base 0.222, GLM-5.3 0.242, DeepSeek-V4-Pro-Base 0.243, Nemotron-550B-Base 0.247. Mirendil says Large and XL both already exceed Kimi-K3 (as summarised by Import AI)
- Mirendil: ‘We don’t think we have reached fully autonomous recursive self-improvement yet, but we are closer than we have ever been.’ Jack Clark’s Import AI 478 (Oct 26) headlined it ‘Mirendil shows what RSI means’
What happened
Mirendil says it pretrained a large coding base model without a pretraining team: one researcher directed AI agents that wrote the training stack, ran the experiments and diagnosed the failures. Its only evidence is a compression metric (bits per byte) on outside code. No downstream benchmark of a finished model, no weights and no technical report have been published.
Why it matters
Pretraining a base model of this class used to take large teams and nine-figure budgets. If the claim holds, agents are already compressing frontier-lab R&D into one person’s work, which is the early stage of AI-driven AI research that labs and forecasters watch for. It is still a self-reported result from a startup that is raising money.
Sources
2 sources from 2 sites. Numbers match the chips in the text.
2 sources: 1 primary, 1 reaction
Primary
- Mirendil: Pretraining at scale without a pretraining teammirendil.com, official
Reactions
- Import AI 478: Math for alignment; ProgramBench; and Mirendil shows what RSI meansimportai.substack.com, discussion
Changes
- Filed (Mirendil post read directly; five days after publication)