AI Weekly Malaysia

Back to items Summaries

DaisyRose-27B: Pushing the Frontier of Post-Training Efficiency

ID
47978
Status
summarized
Published
19 Nov 2026, 3:54 AM
Fetched
19 Nov 2026, 4:06 AM
Provider
Hugging Face Blog
Category
developer-ai
Original URL
https://huggingface.co/blog/RadixArk/daisyrose-27b-pushing-the-frontier-of-post-trainin
Source URL
https://huggingface.co/blog/feed.xml

Summary

Score
6.5
Created
19 Nov 2026, 4:06 AM
Tags
Audience
developersai_ml_learnersai_agent_usersstartup_founders

What happened

RadixArk released DaisyRose-27B, its first model, built on Qwen3.8-27B and trained with reinforcement learning using Miles on 1,200 curated long-horizon agent tasks over 18 days on 144 H200 GPUs, with sessions up to 1M tokens. It keeps Qwen3.8-27B's architecture, tokenizer, chat template, image understanding and 256K context window, and ships Apache 2.0 checkpoints in BF16, FP8 and NVFP4. Reported gains over the base model include Terminal-Bench 4.1 pass@1 rising from 5.6% to 12.0%, DeepSWE v1.1 from 42.2 to 46.2, and SWE Bench Pro v2 from 91.3 to 94.7, though the team notes scores use the higher of the BF16 and FP8 runs wherever both were evaluated.

Why it matters

The interesting number is the training recipe, not the leaderboard: 1,200 tasks and 18 days on 144 H200s is a post-training budget a funded startup could plausibly attempt, and Apache 2.0 weights in FP8/NVFP4 mean you can self-host and benchmark it against your own agent workloads instead of trusting the card. But treat the Terminal-Bench 5.6% to 12.0% jump carefully — doubling a very low pass rate still leaves it failing roughly seven in eight Terminal-Bench 4.1 tasks, and the 'higher of BF16 and FP8' reporting convention means you should re-run the eval on the exact checkpoint and quantisation you plan to deploy before swapping it into an agent pipeline.

Discussion angle

Would you trust a self-reported benchmark where the headline is 5.6% to 12.0% on Terminal-Bench 4.1 and the score is the better of two quantisation runs — and what would your own eval harness need to look like before you'd put this behind an agent in production?

Top