Prime Intellect Releases NanoGPT Speedrun Traces and Setup
Prime Intellect open-sourced traces and setup from 153 autonomous NanoGPT optimizer runs across 18 frontier models.
Prime Intellect released full traces, scratchpads, reasoning streams, and experiment setup from its NanoGPT speedrun. The materials cover 153 autonomous runs by 18 frontier models. Company posts call the effort the largest open test of agents iterating on a research environment. They note models chose which experiments to run, managed benchmark noise, and sometimes built small simulations. One quoted observer said models understood the objects deeply but produced few genuinely new ideas. The data appear on the Prime Intellect research page.
We ran the largest open experiment on how frontier models do AI research. 100+ autonomous runs across 10+ models, sandboxed on 8xH200s for up to 8 days, iterating on the nanoGPT optimizer track. Best runs closed 82% of the gap to a record built by dozens of humans over months.
Reactions
58 postsPI 
Prime Intellect@PrimeIntellectThe task: iterate on a 124M GPT training recipe from a shared baseline, only changing optimizer related hyperparameters, no internet access. We tested Fable 5, Opus 5, GPT-5.6 Sol, Kimi K3, Grok 4.5, GLM 5.2, Muse Spark 1.1, DeepSeek V4 Pro, Grok 4.6, Muse Spark 1.2, Qwen 3.8…
- likes: 153
- replies: 5
- bookmarks: 23
- reposts: 4
PI Prime Intellect@PrimeIntellectWhat separated the strongest models: which experiments to run, how to navigate the benchmark's inherent noise, and which old negatives to revisit as the recipe changed. Some even built small simulations to isolate a mechanism before deciding if another GPU run was worth it.
- likes: 123
- replies: 1
- bookmarks: 16
- reposts: 1
PI Prime Intellect@PrimeIntellectOur Prime Agent harness gives models a persistent IPython kernel, which can help them build their own research workflows. Kimi K3 built tools for controlled optimizer variants, loss-curve comparisons and Newton-Schulz tuning, then revised its hypothesis when its cleaner update…
- likes: 99
- replies: 1
- bookmarks: 9
- reposts: 0
PI Prime Intellect@PrimeIntellectAs research direction, we think multi-agent harnesses can make these experiments much cheaper (and better) by using smaller open models for monitoring and implementation. We also want to extend speedruns to more of the training stack and scale the runs themselves.
- likes: 101
- replies: 2
- bookmarks: 7
- reposts: 1
PI Prime Intellect@PrimeIntellectWe release everything: full traces, scratchpads, reasoning streams from open-weight models, and our experiment setup. Explore the results: https://primeintellect.ai/research/nanogpt-speedrun
- likes: 135
- replies: 2
- bookmarks: 46
- reposts: 2
EL elie@eliebakouchto my knowledge this is the largest open experiment on autonomous agents iterating on a research environment we scaled runtime, compute, diversity of models and harnesses. as a comparison, similar tasks on oai/anthropic system cards are anthropic "optimizing an llm training on…
- likes: 768
- replies: 47
- bookmarks: 418
- reposts: 70
