How much can you improve a small model in 18 days on 144 GPUs?
Meet DaisyRose-27B, RadixArk's first open-weight model! It is trained with RL on Qwen3.8-27B, with 1,200 curated long-horizon agent tasks for 18 days on 144 H200s. Powered by Miles and our own training backend, we stretched agent sessions up to 1M tokens and raised the Terminal-Bench 4.1 score from 5.6% to 12.0%.
See how we pull it off ⬇️
Other TPU folks at Google just published a great public guide on TPUs:
If you’re curious about TPUs and hardware-friendly model co-design this is a great resource to check out!
Attaching some graphics from the docs, all work is by the authors(not me)
https://t.co/c2trC0yEtO
Introducing SID-2, a new search model built on compute-scaling swarms.
Our swarms find better results than Opus and Astra, while being 8x faster and 20x lower cost.
By 2030, AI will write an internet's worth of text every 24 hours. Current search pipelines were built on hand-designed rules, adding more won't scale to meet this incoming volume of data.
SID-2 is available on our API today. Try it on your most demanding searches.
@liyanc98@andrew_n_carr What shapes did you optimize for and which you didn’t? Curious what is the line of sufficient shape size to saturate Tensor Cores
Training with fewer bits is faster, but quantization error can interfere with learning.
Today on Connectionism, we detail how we trained Inkling 2 Small with a native 4-bit forward pass from the first step of pretraining through the final RL rollout.
https://t.co/CIMECSEpD0
2/N In async RL, sampler q generates while trainer p keeps updating. Once q≠p, two errors appear:
• Drift: a push toward/away from q that has nothing to do with which token is better.
• Reweighting: tokens arrive at q’s frequencies, not p’s.
SC centers the score to zero mean under q, removing drift but not reweighting error.
Importance sampling (IS) corrects both but risks extreme weights. Truncation (TIS) caps these weights but leaves residual drift, which SC can remove.
We identify a blind spot in (M)OPD for post-RL consolidation:
Students may not inherit their teachers’ generalization capabilities, though they achieve near-perfect in-distribution distillation performance. 🧵
Wow NVFP4 pre-training and post-training (sampling + training), StableReinforce is working very well I guess
Holy 17 step policy staleness lmao
5:3 sampler to trainer ratio
2048 GPUs for 9 days, so RL training costs about $1.2M
Introducing Inkling 2 Small. Scoring 31 on the AA Intelligence Index, it's a big jump over its predecessor.
Inkling 2 Small is available for fine-tuning and inference on Tinker today, with full weight release planned soon.
https://t.co/VR02tw73fJ