Vibeleaderboard
← All Intel
Intel / article

DaisyRose-27B

Source
radixark.com
Author
de_sty
Date

Context

DaisyRose-27B shows that a small, carefully chosen set of tasks can noticeably lift a mid-sized open model on long multi-step work, and its release is just as much a showcase for the infrastructure that trained it. RadixArk, the company behind the open-source reinforcement learning framework Miles, took the Qwen team's Qwen3.8-27B and trained it further with reinforcement learning, in which the model attempts tasks, is graded, and is updated toward the attempts that succeed. According to RadixArk's announcement, the run used about 1,200 curated tasks for 18 days on 144 H200 GPUs and raised the model's Terminal Bench 4.1 pass@1 (the share of tasks solved on a single attempt, averaged here over 8 runs per task) from 5.6% to 12.0%. The weights can be downloaded in BF16, FP8 and NVFP4 versions, which store the same model at progressively lower numerical precision so it needs less GPU memory. The Hugging Face version of the post adds that they are released under the Apache 2.0 license and keep the base model's architecture, tokenizer and 256K-token .

RadixArk attributes the gain mainly to task design. Working with Eigent AI, it built thousands of candidate tasks in machine learning, software engineering, science and other fields and kept about 1,200. Each kept task was redesigned around a mistake the base model actually made, tuned so that its solve rate was neither too high nor too low for the model to learn from, and built as a chain of dependent problems: one attempt could run for up to 1.5 hours in a , with more than 100 tool calls. The team says it also reviewed every task alongside the model's attempts and fixed or removed any that could be passed without really solving it. That guards against reward hacking, where a model learns to satisfy the grader instead of doing the task.

Outside Terminal Bench 4.1, which RadixArk calls a notoriously hard long-horizon benchmark, the improvements are small. In RadixArk's own comparison table, DaisyRose-27B beats its base model on all seven benchmarks listed, but on the other six the margin ranges from 0.4 points (HLE, text only) to 4.0 points (DeepSWE v1.1). Its 12.0% on Terminal Bench 4.1 is also below several models in the same table, including MiMo V2.6 Flash at 16.1%, GPT-6 Luna at 15.2% and Inkling 2 Small at 13.0%, although MiMo V2.6 Flash and Inkling 2 Small are listed with far more total parameters (309 billion and 594 billion) but fewer activated parameters (15 billion and 17 billion, against DaisyRose's 27 billion). Every figure comes from RadixArk's post. DaisyRose's Terminal Bench 4.1 score was measured with the company's internal , and the post does not say where the other models' scores came from. The post cites no independent evaluation.

The rest of the post is about cost, and its strongest numbers describe RadixArk's internal systems rather than what it released. Training ran fully asynchronously: 128 GPUs kept running the agent on tasks while 16 trained on the finished attempts, and updated weights went back to the agent GPUs after each step. RadixArk says its internal training backend kept 37% of each GPU's peak compute doing useful training work, a measure called model FLOPs utilization, against 16% for the open-source Megatron-based trainer on comparable runs, so the same training needs 57% less GPU time. It also says its sandbox service, built on the open-source AgentENV project, cost about 80% less than running the same sandboxes with unnamed popular cloud sandbox providers. Miles is open source, but the post calls the training backend internal, does not say the sandbox service is available, and closes by offering the approach to teams building specialized models. Anyone can download and test the model; the cost comparisons rest on RadixArk's word.

Terms in this piece · Glossary
  • AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
  • context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.
  • sandbox — An isolated environment where AI-generated code or agent actions run without being able to touch anything real.
  • agent harness — The scaffolding around a model that turns it into a working agent — the loop, the tools it can call, and the rules for when to stop.
Read the source www.radixark.com
Recommended reads
Comments

Checking sign-in…

Loading comments…