Tasks and inputs
DaisyRose-27B is listed for Image-text-to-text, Text generation, Coding & Code Generation. It accepts text, image and produces text. Use the publisher’s input format and preprocessing for your chosen deployment.
SyncDev model directory · Profile updated Nov 21, 2026
DaisyRose-27B is RadixArk's Apache 2.0, RL-only post-train of Qwen3.8-27B for agentic coding, with a 256K context and image input.
The specification score summarizes recorded features and limits. It is not a benchmark result.
Pricing conditions: Open weights; RadixArk announces no hosted API or pricing.
Recorded modality support shows what this model can accept and produce. An undocumented modality is shown as unknown rather than assumed to be unsupported.
Accepts
Produces
Eighteen days, 144 H200 GPUs, roughly 63,000 GPU-hours. That is the entire post-training budget behind DaisyRose-27B, and it is the real story here. RadixArk took an existing open model, Qwen3.8-27B, and improved it using reinforcement learning alone: no supervised fine-tuning, no distillation from a bigger teacher. The result went out with open weights on 18 November 2026 under Apache 2.0.
What you get is the same model shape with sharper agent behaviour. Architecture, tokenizer, chat template, image understanding and the 262,144-token context are all inherited unchanged, which makes it a drop-in replacement in any stack that already serves the base model.
Every figure below is publisher-reported and compares DaisyRose-27B with its own base.
| Benchmark | DaisyRose-27B | Qwen3.8-27B |
|---|---|---|
| Terminal Bench 4.1 | 12.0 | 5.6 |
| Terminal Bench 2.1 | 75.5 | 73.0 |
| DeepSWE v1.1 | 46.2 | 42.2 |
| SWE Bench Pro v2 | 94.7 | 91.3 |
| HLE (text only) | 34.3 | 33.9 |
| IFBench | 82.4 | 79.5 |
| GPQA Diamond | 90.7 | 89.2 |
Read that honestly. Terminal Bench 4.1 more than doubles, which is the headline. The rest are gains of a few points, and HLE barely moves. In RadixArk's own wider comparison, much larger models such as Inkling 2 Small still beat it on most rows; it leads only on IFBench. This is a better 27B model, not a giant-killer.
About 1,200 long-horizon agent tasks made the final cut, each designed around a mistake the base model makes and tuned so it is neither trivially easy nor hopeless. A single attempt can run up to 1.5 hours in a sandbox, with more than 100 tool calls and over 256K output tokens. Some training sessions stretched to a million tokens.
That last number confuses people. It describes training sessions, not the inference context window, which stays at 256K. If you see a 1M context quoted elsewhere, it is a misreading.
The run was fully asynchronous on RadixArk's open-source Miles framework: 128 GPUs generated agent trajectories while 16 trained on the finished ones, pushing fresh weights back after every step.
Leave max_tokens unset on hard problems. Thinking can run to tens of thousands of tokens, and a cap will truncate the answer rather than shorten the reasoning.
If you already serve Qwen3.8-27B for agentic coding or terminal work, the test is cheap: point the same harness at the new weights and rerun your own failing tasks. Watch two things. Does it finish more multi-step jobs? And does the longer default reasoning blow your latency budget? There is no hosted API from RadixArk, so self-hosting is the only route.
For other open models in this size class, browse code generation models or image-text-to-text models.
Compare the model on the work you need to complete.
DaisyRose-27B is listed for Image-text-to-text, Text generation, Coding & Code Generation. It accepts text, image and produces text. Use the publisher’s input format and preprocessing for your chosen deployment.
Documented limits: 262.14K context. Measure a completed task using the prompts, documents and tools your application will use. Include reasoning and retries in the result.
The weights are downloadable under Apache-2.0. Check runtime support and size the complete model and working context for your expected concurrency.
These records share a directory category. Check the source card, output type and deployment requirements before treating a model as a replacement.
Thinkingmachines
Inkling 2 Mini is Thinking Machines' Apache 2.0 looped MoE model (26.4B total, 3.4B active) for agentic coding that runs on a 32 GB laptop.
Microsoft
MAI Code 2 is Microsoft AI's coding model for multi-file work, debugging and frontend builds, live in GitHub Copilot at $0.50/$2.50 per 1M (promo).
Microsoft
MAI Code 2 Flash is Microsoft AI's low-cost coding model for edits, fixes and explanations, on every GitHub Copilot plan including Free and Student.
Microsoft
MAI Cyber 2 Flash is Microsoft AI's security model for vulnerability discovery and threat investigation, with a 1M-token context, shipping in MDASH.
Gemini 4 Hydrogen is Google's fast, lower-cost Gemini 4 model for everyday coding and agentic execution, with a 1M context, xhigh thinking and a free tier.
OpenAI
GPT-6 Cyber is OpenAI's restricted security model for authorised vulnerability investigation, reproduction and fix verification, for Daybreak Red customers.