Understand this story
On October 22, 2026, Mirendil released the code base model Mirendil-Base. According to earlier reports, the model was pretrained from scratch in 3.5 months by a team of Agents led by a single researcher, with a parameter scale of 100 billion to 500 billion, at a cost of less than 10 million USD, using only public data, without distillation, without using synthetic data, and without purchasing data, with more than 99% of the code written by Agents. On code prediction loss, it outperforms GLM-5.3, DeepSeek-V4-Pro-Base and Nemotron-550B, but falls short of Kimi-K3; the model does not open its weights, has no API, and there is not yet evaluation on tasks such as SWE-bench. Subsequent reports said that in a cross-entropy test on new code repositories, Kimi K3 ranked first with 0.193, Mirendil-Base ranked second with 0.222, followed by GLM-5.3, DeepSeek-V4 and Nemotron; the reports also noted that base models typically outperform post-trained models on this test, so the actual gap between Mirendil-Base and Kimi K3 is larger than the scores suggest.
AI-generated from public reports
Latest development · Report
Mirendil and Recursive release RSI results, NVIDIA researcher discusses survival rules for researchersDevelopments
1 confirmed developments
- Source:TMTPost · AGI
Reports and discussion timeline
Browse reports and discussion by time. Their evidence levels remain distinct.
Oct 23
36Kr · AI
Mirendil and Recursive release RSI results, NVIDIA researcher discusses survival rules for researchers
Startup Mirendil claims that one researcher, with the help of AI agents, wrote over 99% of the code and trained a 100B–500B parameter coding foundation model, Mirendil-Base, in 3.5 months with under $10 million in compute; Tian Yuandong's startup Recursive released Flint, an open-weight 40B coding model, backed by the Eureka Machine system that automatically generates ideas, implements, tests, and decides the next step. NVIDIA senior researcher Ali Hatamizadeh wrote about this, arguing that conventional pretraining recipes are now mature and public, and that researchers' value has shifted to deciding which experiments to run and judging which results are trustworthy; he lists seven open topics: few-shot efficient learning, new architectures, continual learning, long-horizon reasoning reliability, eliminating hallucinations, new evaluation systems, and interpretability and control.
TMTPost · AGI
Mirendil releases a hundred-billion-parameter coding foundation model Mirendil-Base, trained by 1 researcher and an Agent
Mirendil releases a coding-specific foundation model Mirendil-Base, with a parameter range of 100 billion to 500 billion, completed by 1 researcher and an Agent in three and a half months, at a cost of less than 10 million USD, with over 99% of the code written by the Agent. The official statement says training used entirely public data such as GitHub open-source code, public web pages, and public datasets, with no purchased data, no distillation, and no synthetic data. In the cross-entropy test on new code repositories, Kimi K3 ranks first with 0.193, Mirendil-Base ranks second with 0.222, followed by GLM-5.3, DeepSeek-V4, and Nemotron; the article also points out that foundation models are usually better than post-trained models in this test, so the actual gap between Mirendil-Base and Kimi K3 is larger than the scores show.
TMTPost · AGIRelated discussion
TMTPost · AGI joined this discussion
This post contributes to heat; it is not a formal report or confirmed fact.
Oct 22
indigo
Mirendil Releases Mirendil-Base Code Foundation Model, Pretraining Completed in 3.5 Months
Mirendil has released its code foundation model Mirendil-Base, pretrained from scratch by one researcher leading a group of Agents over 3.5 months at a cost of under $10 million, producing a model with 100–500 billion parameters. It used only public data—no distillation, no synthetic data, and no purchased data—and more than 99% of the code was written by Agents. On code prediction loss it outperforms GLM-5.3, DeepSeek-V4-Pro-Base, and Nemotron-550B, but falls short of Kimi-K3. The model does not release its weights, has no API, and has no task evaluations such as SWE-bench yet; the author views this as more of an efficiency win for a senior researcher plus Agents, and the method relies on third-party Coding Agents and is easy to replicate.
Heat trend for this story
- Current
- 8
- Period peak
- 14 Oct 23 18:00
- 24-hour change
- —
Move the pointer or select the chart to inspect hourly heat. Use the left and right arrow keys with a keyboard.