One researcher pretrains coding model Mirendil-Base in 3.5 months for under $10M
bneyshabur · x · 2026-10-22
Startup Mirendil says a single researcher pretrained its first coding base model, Mirendil-Base, in 3.5 months with under $10M of compute, including all research and the final run — no pretraining team needed.
- Per the shared chart (cross-entropy bits/byte on held-out code, lower is better), Mirendil-Base hits 0.2803, beating Nemotron-550B-Base (0.2918), GLM-5.3 (0.2926), and DeepSeek-V4-Pro-Base (0.3013), though trailing Kimi-K3 (0.2387)
- Larger models are scaling well: Base-XL already at 0.2181 (19% trained) and Base-Large at 0.2177 (80% trained)
- The team frames it as a step toward one AI-empowered researcher doing what once took an entire research org
More from Models
- YC F26 startup Latent reads model internals to catch hallucinations at 530x lower cost — ycombinator · 2026-10-22
- Perplexity's JOLT matches GRPO with 6.5x fewer training tokens, JOLT+ wins on all five benchmarks — perplexity_ai · 2026-10-22
- Recursive AI ships Flint, a 40B-A2.5B MoE claimed to match Qwen3.6-27B, weights on Hugging Face — ChengleiSi · 2026-10-22
- DeepSeek ships V4.1-Flash-1017, now the #1 open model on RSI-Inference Bench — scaling01 · 2026-10-22
- OpenAI's Cybersecurity Guardrails Block a Dev From Auditing His Own Patched Server — Next_Technology6361 · 2026-10-22
- TPN-006 hits Hugging Face: 7.7GB compressed model beats its 17GB Mistral teacher — markjeffrey · 2026-10-22