Open DaisyRose-27B solves 12% of Terminal-Bench tasks, up from 5.6%
RadixArk has released DaisyRose-27B, an open model that spent 18 days being fine-tuned on 144 H200 GPUs to work in the terminal like a developer. On Terminal-Bench 4.1, the share of tasks solved on the first attempt rose from 5.6% to 12.0% (GPT-6 Luna scores 15.2%). Each of the 1 200 training tasks targeted a specific failure of the base Qwen3.8-27B. The takeaway: an agent's frequent failures can be turned into a training set, and Apache 2.0 lets you ship the model in a product.
AI-processed from X: @Chinazhidx; edited by Hamidun News
Imagine an agent told to fix a broken build in the terminal: find the error in the logs, patch the config, rerun the tests, and not break the neighboring service. Most open models still stall on tasks like this. The startup RadixArk has released its first open-weights model, DaisyRose-27B, and doubled its Terminal-Bench 4.
1 result: the share of tasks the agent solves on the first attempt (the pass@1 metric) rose from 5.6% to 12.0%.
For comparison, GPT-6 Luna scores 15.2%. The key idea is that the model was trained not on a general data mix but on tasks that each targeted a specific failure of the base model.
For developers, this is a practical signal: an agent's weak spots can be turned into training material.
An experiment that turned into a product
According to co-founder and CTO Banghua Zhu, the project did not start as a model release but as a stress test of the infrastructure. The team wanted to find out whether its system could handle training agents on a context of a million tokens, whether GPU utilization could be raised, and whether the sandboxes where agents run code could be made cheaper. The results beat expectations, and the model grew out of the stress test.
The base was Qwen3.8-27B. The team fine-tuned it with reinforcement learning through the Miles framework: the model tries to solve a task, receives a score, and adjusts. Together with Eigent AI, the team assembled 1 200 carefully selected tasks. It's like studying not from a general textbook but from a list of your own mistakes, with a tutor picking exercises for exactly those.
The tasks were long. A single attempt could run up to 1.5 hours, involve more than 100 tool calls, and produce more than 256 thousand tokens. The entire training took 18 days on 144 NVIDIA H200 GPUs.
What the numbers show
The most striking achievement is on the infrastructure side. According to RadixArk, its in-house training backend reached 37% MFU, meaning it used 37% of the hardware's theoretical maximum. A comparable open-source trainer built on Megatron reached 16%. Throughput was 16 800 tokens per second versus 8 300 on the same 144 GPUs. As a result, the same workload needs 57% less GPU time. For anyone paying for H200 rental by the hour, that is a direct saving.
On quality, DaisyRose-27B beat the base Qwen3.8-27B on all seven published benchmarks. The biggest gain is on Terminal-Bench 4.1: more than double, from 5.6% to 12.0%. The other gains are more modest:
- DeepSWE v1.1: from 42.2 to 46.2;
- SWE-Bench Pro v2: from 91.3 to 94.7;
- GPQA Diamond: from 89.2 to 90.7.
Notably, there is a gain even on GPQA Diamond, a test of scientific questions, even though training used agentic tasks.
Why this matters for developers and founders
First takeaway: an agent's quality can be raised with a small set of tasks if each one targets a real weakness of the model. Twelve hundred tasks is not internet scale, but a volume a small team with a partner can assemble. If you have logs where your agent consistently fails, you can turn them into your own training tasks.
Second takeaway: the bottleneck in long agentic runs is often the training stack, not the model itself. When the same work takes 57% less GPU time, iterations get cheaper and you can test more hypotheses.
Third: the Apache 2.0 license permits commercial use, and 27 billion parameters keep the model within reach hardware-wise. According to the authors, it runs on a single RTX 5090 or on an NVIDIA DGX Spark, so you can test the agent locally without renting a cluster.
Caveats
All the figures have so far been published by RadixArk itself, and there are no independent runs. The comparison with the Megatron-based trainer was made on the company's own workload, and the gap may differ on other tasks. A 12.0% result on Terminal-Bench still means the vast majority of tasks are not solved on the first attempt, so such an agent is not yet ready for production without human oversight. It is also unknown how the model will behave on tasks unlike the 1 200 selected ones. Still, as an example of how targeted fine-tuning and efficient infrastructure deliver a noticeable gain on a modest budget, DaisyRose-27B deserves attention.
- Nov 19, 2026The same Gemma 4 wrote a backend that passed 80% of checks in opencode, but only 41% in OpenHands
- Nov 19, 2026OpenAI's coding agents would cost $2400 a day per researcher at API prices
- Nov 19, 2026AWS AgentCore Agents Can Now Reach a Private Keycloak Behind a Corporate CA
- Nov 19, 2026Smysl.AI: An Agentic RAG Over Company Documents, Built in Kotlin Without Python
Key Questions Answered
What are the key details on An experiment that turned into a product?
According to co-founder and CTO Banghua Zhu, the project did not start as a model release but as a stress test of the infrastructure. The team wanted to find out whether its system could handle training agents on a context of a million tokens, whether GPU utilization could be raised, and whether the sandboxes where agents run code could be made cheaper. The results beat expectations, and the model grew out of the stress test. The base was Qwen3.8-27B.
What the numbers show?
The most striking achievement is on the infrastructure side. According to RadixArk, its in-house training backend reached 37% MFU, meaning it used 37% of the hardware's theoretical maximum. A comparable open-source trainer built on Megatron reached 16%. Throughput was 16 800 tokens per second versus 8 300 on the same 144 GPUs. As a result, the same workload needs 57% less GPU time. For anyone paying for H200 rental by the hour, that is a direct saving.
Why this matters for developers and founders?
First takeaway: an agent's quality can be raised with a small set of tasks if each one targets a real weakness of the model. Twelve hundred tasks is not internet scale, but a volume a small team with a partner can assemble. If you have logs where your agent consistently fails, you can turn them into your own training tasks. Second takeaway: the bottleneck in long agentic runs is often the training stack, not the model itself.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.