Image + textComputer-use agentsCodeMath and reasoning

DaisyRose-27B

The first RadixArk model: a multimodal reasoning model based on Qwen3.8-27B, trained with reinforcement learning only, without fine-tuning on examples and without distillation. It was trained on about 1,200 selected long agentic tasks with sessions of up to a million tokens, for 18 days on 144 H200 graphics cards. By the authors measurements, on Terminal Bench 4.1 the result rose from 5.6 to 12.0, and it beats the base model on most other tests of code, reasoning and instructions. It keeps the architecture, tokenizer, image understanding and 256 thousand token context of the base model.

Developer
RadixArk, not stated
First release
Nov 2026
Latest release
Nov 2026
Sizes
27B
License
Commercial use allowedApache 2.0 (based on Qwen3.8-27B)
Russian
Not stated
Running
On your own serverNeeds a GPU
Industries
Software development, Customer support

What it does

  • an agent for programming and the terminal on your own graphics card
  • long tasks with tool calls
  • a base for your own reinforcement learning

Where it is used

software developmentresearch groupscompanies with their own servers

Hardware requirements

LaptopLaptop or regular PC, up to 8 GB of VRAM — smaller versions
no versions
1 GPUOne GPU with 16–80 GB — mid-size versions
fits
ClusterServer with several GPUs — flagship versions
no versions

Versions

  1. DaisyRose-27B (BF16)

How to run it

On your own serverWeights are downloaded from Hugging Face and served with vLLM (load and API) or llama.cpp (modest hardware). The model then runs in a closed network with no per-request fees.Hugging Face
How much hardware you needCalculate the VRAM for the model size, context length and number of concurrent requests.Open the hardware calculator

I can set this up end to end: pick the model size, deploy it on your server and connect it to your systems.

Frequently asked questions

Can DaisyRose-27B be used in a commercial project?

Yes. License: Apache 2.0 (based on Qwen3.8-27B). It allows commercial use, but it is still worth having a lawyer review the license before launch.

What hardware does DaisyRose-27B need?

At minimum: One GPU with 16–80 GB — mid-size versions. Without a GPU the model is not practical. You can calculate the exact VRAM for your model size and context in the hardware calculator.

Does DaisyRose-27B support Russian?

The model card does not list languages, so Russian support cannot be promised — it has to be tested on your own examples.

Where can I download DaisyRose-27B and what does it cost?

The DaisyRose-27B weights are open and free to download. You only pay for the hardware it runs on and for the setup. Source links are at the bottom of this page.

How I deploy it for clients

  1. SelectionI pick the model size for your task and hardware and test it on your examples.
  2. DeploymentI deploy it on your server or in a closed network and provide an API.
  3. Fine-tuningI fine-tune it on your data (LoRA) or connect a knowledge base — whichever is cheaper for the task.
  4. IntegrationI connect it to your CRM, ERP, bot, website or team chat and set up monitoring.

Similar models

TextQwenAlibaba · ChinaCommercial use with conditions

A family of language models with strong Russian language support, from small versions for a laptop to a flagship on par with commercial APIs.

Details
TextArete-27BNous Research · USACommercial use allowed

A fine-tune of the Chinese Qwen3.8-27B from Nous Research whose aim is to remove evasive, refusing and officially-Chinese-sounding answers on politically sensitive topics. By the authors own estimate on their test set, the factuality score rose from 2.32 to 2.98 out of 5 and the share of answers with propaganda fell from 22% to 0.9%, with general and agentic skills not visibly dropping. The authors state plainly that this transfers Western values onto a Chinese model and note that they addressed a cause but not all the behavior. The Propaganda_Eval test set is published too. The model reflects the authors choice of values, so for tasks that need a neutral tone it should be checked on your own examples.

Details
TextNanbeige 4.3Nanbeige · ChinaCommercial use allowed

A 33B model that the authors call the senior teacher for future lightweight agents: they plan to train small models (for example 3B) on it. It is strong in programming and office tasks: on the JobBench test across 35 professions it beats the baseline in most, according to the authors. Context up to 512 thousand tokens, reasoning can be switched on and off, and tool calling works best in the XML format. The authors compare it with Qwen3.8-27B, Muse-Glimmer-30B and Gemma 4 31B and say it leads on most tests; the figures are their own. A base version is published too.

Details

Source: huggingface.co/RadixArk/DaisyRose-27B. Data checked against the model card on 22 Sep 2026. Have a lawyer review the license before commercial launch.