TAU-HOME.COM
LOADING

RL-Only Coding Model DaisyRose-27B Released and llama.cpp Adopts Semantic Versioning with v1.0.0

RadixArk has released DaisyRose-27B, an open-source coding model trained purely through reinforcement learning without supervised fine-tuning or distillation, a

tau · November 21, 2026

#DaisyRose-27B #llama.cpp #Qwen #ReinforcementLearning #LocalAI

RL-Only Coding Model DaisyRose-27B Released and llama.cpp Adopts Semantic Versioning with v1.0.0

The third week of November 2026 brought significant developments to the open-source and local artificial intelligence ecosystem. Research and development firm RadixArk has released 'DaisyRose-27B', an open-source coding model trained exclusively through reinforcement learning (RL) without utilizing supervised fine-tuning (SFT) or knowledge distillation. Concurrently, the core libraries powering local model execution, 'llama.cpp' and 'ggml', officially adopted semantic versioning starting with v1.0.0 to manage application programming interface (API) and application binary interface (ABI) stability.

Conceptual diagram illustrating local AI ecosystem updates including the reinforcement learning model DaisyRose-27B and llama.cpp v1.0.0

Image source: https://toolhalla.ai/blog/this-week-in-local-ai-2026-11-20

These developments highlight how the local AI landscape is maturing beyond experimental builds into practical stability and advanced training methodologies. In particular, the attempt to boost specialized coding agent performance using pure reinforcement learning without supervised data, paired with the structural versioning shift for ecosystem dependency management, marks a crucial milestone for developers and local AI practitioners alike.

DaisyRose-27B: Pure Reinforcement Learning Without Supervised Fine-Tuning

Released by RadixArk on November 18, 2026, 'DaisyRose-27B' is an Apache 2.0 open-source model built upon Alibaba's 'Qwen3.8-27B' base model. The creators trained the model using an internal reinforcement learning environment called Miles.

The defining characteristic of DaisyRose-27B is that its fine-tuning process relied entirely on reinforcement learning, bypassing traditional supervised fine-tuning (where models are trained directly on correct answer datasets) and distillation entirely. According to RadixArk's internal evaluations, its performance on 'Terminal Bench 4.1 pass@1'—a benchmark measuring terminal coding agent capabilities—rose from 5.6% to 12.0%. However, these figures are the authors' own measurements averaged over 8 runs per task using an internal mini-SWE-agent harness, and real-world performance may vary depending on execution environments.

Because DaisyRose-27B preserves the base model's architecture, tokenizer (the mechanism that breaks text into tokens for AI processing), and chat template, RadixArk notes that it serves as a drop-in replacement for Qwen3.8-27B without requiring structural modifications.

Hardware Requirements and Quantization Formats for Local Deployment

For users seeking to run the model locally, community developers such as bartowski have provided GGUF (GPT-Generated Unified Format, a compressed file format optimized for running large language models on consumer CPU and hardware) quantization files.

  • BF16 / FP8 / NVFP4 Checkpoints: RadixArk has published official precision checkpoints alongside the model weights.
  • Quantization File Sizes: Community GGUF files range from Q4_K_M (4-bit quantization) at approximately 16.14 GiB, IQ4_XS (ultra-low loss 4-bit) at 14.32 GiB, to Q3_K_M (3-bit quantization) at 12.41 GiB.

When deploying the Q4_K_M file on a standard GPU or laptop configuration with 16 GB of video memory, the model file consumes a large portion of available capacity before accounting for context memory. Users should consider starting with IQ4_XS alongside a shorter context window or opting for lighter 3-bit variants like Q3_K_M.

llama.cpp and ggml Libraries Adopt Semantic Versioning (v1.0.0)

During the same week, the foundational libraries enabling local model execution—'llama.cpp' and 'ggml'—reached a major structural milestone. The development team officially announced on November 13, 2026, that both projects are transitioning from sequential build numbers to formal semantic versioning (v1.0.0).

Going forward, breaking changes to the public C application programming interface (API) will trigger a major version bump. Minor versions will cover new features, backend updates, and model architecture support, while patch versions will handle bug fixes.

  • Enhanced Stability: Given that numerous downstream applications and user interfaces depend on llama.cpp, this versioning policy provides a predictable and stable contract for developers and distributors.
  • Ongoing Build Updates: Sequential numbered builds continue alongside semantic releases; for instance, build b12406 introduced Inkling 2 support for combined text and vision workloads, reflecting rapid active development.

Sources

Impact-Site-Verification: 9b3ebd10-6c45-44cf-a014-5b8c19b7b285