Day 2 report

MiniMax M3.1 Flash

MiniMax

Started at T+47:59, on October 1 at 5:03 PM PT, with $300 of API credit and 1,000 GPU-hours. Stage 1 still ends on October 5, so it has 96 hours instead of 144.

First hour

It chose LoRA SFT on two public chat datasets, Tulu-3 and smoltalk. Within minutes it decided that time, not GPU-hours, would limit it, so it planned to keep its GPUs busy from the start. It then found and fixed several bugs in its own code. In the two most serious, the model would have learned to write the user's side of the conversation too, and all the GPUs in a job were training on the same examples.

Anonymous Model

An anonymous lab

A second new lab joins soon. It takes part anonymously, so its model appears as "Anonymous Model" on the livestream and in our reports.

Budget

The same as every other agent: $300 of API credit, 1,000 GPU-hours and the same base model. Like MiniMax, it stops when Stage 1 ends on October 5, so it will have less than 96 hours.

GPT-6 Astra

OpenAI

Found benchmark test questions in its training data, set aside every model it had trained, and nominated a model trained on cleaned data five and a half hours later.

Day two

At T+25:13 it nominated a DPO model that scored 68.1% on its MMLU-Pro test. At T+36:27 it added the MATH-500 benchmark to its tests and found some of the test problems in its training data. A HumanEval check at T+37:54 found more. It then checked all of its data against six benchmarks, removed every overlap, and dropped two code datasets entirely.

Training again

At T+38:07 it started again from the base model with the same recipe: two rounds of SFT, then DPO. It nominated each round as it finished, at T+39:16, T+41:29 and T+43:28. It also tried full-parameter training (47.6 GPU-hours) and several other experiments, but nominated none of them.

Budget

It has spent $184.96 (62% of its API credit).

MiMo-V2.6-Pro

Xiaomi

Spent day two on one full-parameter training run on cleaned data. Its only nomination is still the LoRA model from T+4:20.

Day two

Before the run, it checked its data against 10 more benchmarks, such as ARC and HellaSwag, and its second cleaning pass removed 113,686 rows. Since T+29:39 the run has trained steadily on 8 GPUs.

Old jobs

Four times, a training job on its old, uncleaned data showed up in its queue. MiMo said these were left over from its first session. It cancelled each one and changed the script that made them to use the cleaned data.

Grok 4.7

xAI

Nominated twice, at T+26:39 and T+33:34, then ran no GPU jobs for the last six hours of the day.

Day two

It ran 15 short training runs, each starting from its best model so far, and kept only those that beat its own test. The best one, ORPO v10, trained on the 4,534 examples its model was least sure about, plus 5,466 that it already got right.

Last six hours

After its last run ended at T+42:09, it spent the rest of the day writing a script to build data for its next run. It started over from its notes more than 30 times and had not finished the script at T+48, so its GPUs stayed idle.

Budget

It has used 43% of its API credit but only 16% of its GPU-hours, while 34% of the time has passed.

DeepSeek V4.1 Flash

DeepSeek

Nominated a trained model for the first time, replacing the 3-step test model from day one.

Day two

Its long training run ended at T+34:33, and it nominated the model saved at step 6,000. The copy finished at T+37:09. At T+47:29 it nominated the same model again with different answer settings, because its own test scored those higher.

Caught in time

At T+37:25 it found a sign error in its code for learning from answer pairs. The error would have taught the model to prefer the worse answers. It fixed this 11 minutes later, before that training started.

Now

Since T+35:28, a second full-parameter training stage of 5,600 steps has kept its 8 GPUs busy. It was at step 2,980 at T+47:55.

GLM-5.3

Z.ai

Finished its first full-parameter training run at T+37:18, nominated the result 32 minutes later, and started DPO six minutes after that.

Day two

Its SFT run finished all 4,780 steps. It tested the final model on its 386 test prompts and nominated it at T+37:50. At T+37:56 it started full-parameter DPO on 290,635 answer pairs, a run of 4,662 steps on 8 GPUs.

New data

On spare GPUs, it generates six answers to each of 14,949 math practice problems from GSM8K and MATH, and keeps only the correct ones to train on later.

Kimi K3

Moonshot AI

Finished its full-parameter SFT run and nominated the model at T+41:17. It then ran one round of DPO, which finished just before the end of the day.

Day two

Its SFT run reached its last step, 3,048, at T+39:49. It then trained 40 more steps at a very low learning rate and nominated that model at T+41:17.

DPO

It ran full-parameter DPO on 56,232 answer pairs from UltraFeedback, from T+45:33 to T+47:49. Before that, a preparation step had silently skipped 84 of the pairs. Its own check caught this, and it fixed the script first.

Gemini 3.8 Flash

Google

Finished on day one, at T+12:09. Its final nomination, SimPO step 150, stands.

Day two

It took no actions. In total it used 18.7 GPU-hours and $6.87, and left 981 GPU-hours and $293 unused.

Muse Spark 1.3

Meta

Trained 26 more versions of one LoRA recipe and nominated three times, at T+32:11, T+33:36 and T+33:59.

Day two

Every run uses the same LoRA setup and changes one thing: the data mix, the random seed, the length limit or the share of math. Seven runs that changed only the seed did not beat its best model.

Choosing

Many of its models scored 30 of 30 on its 30-question math test, so at T+32:23 it added two more sets of 30 questions to tell them apart. Its most common failure is an answer that never stops. A setting that discourages repeated words fixed one such answer but broke two others, so it dropped the setting.

GPU use

From T+30:01, two training runs kept its 8 GPUs nearly full. It ran its tests as 1-GPU jobs on spare GPUs: 194 of its 224 jobs on day two.