Key takeaways
- RadixArk announced DaisyRose-27B on November 18, 2026. OpenRouter’s exact route radixark/daisyrose-27b was created on November 21 UTC.
- The model accepts text and images and produces text, with a 262,144-token context. Its card lists 27.8 billion total parameters and Apache 2.0 weights.
- OpenRouter lists DeepInfra at $0.15 per million input tokens, $1.875 per million output tokens and $0.0375 per million cached input tokens. This is not a free route.
- RadixArk reports coding gains after reinforcement learning on long-horizon tasks. These are vendor results, not AccessAllGPT measurements or a controlled comparison across all listed models.
What is available now
RadixArk’s DaisyRose-27B has a newly listed OpenRouter API route as of November 21, 2026: radixark/daisyrose-27b. The registry lists a 262,144-token context, text and image inputs, and text output. At retrieval, its exact-route provider endpoint named DeepInfra and quoted $0.15 per million input tokens, $1.875 per million output tokens and $0.0375 per million cached input tokens.
This is a new access path for an existing release, not a model first announced today. RadixArk’s own launch post is dated November 18, 2026. The OpenRouter record’s creation timestamp resolves to November 21 at 00:49:48 UTC. Its date-bearing canonical slug is not the announcement date; builders should request the listed ID rather than constructing an ID from a date.
RadixArk also publishes weights on Hugging Face. We directly verified an ungated public BF16 repository, its Apache 2.0 license declaration and a file inventory containing 18 safetensors shards. The card links FP8, NVFP4 and GGUF variants; we did not download or validate those artifacts. A public listing and provider record establish discoverability, not a successful authenticated inference call or guaranteed access for every account.
A smaller dense model trained for longer agent tasks
DaisyRose is built on Qwen3.8-27B. RadixArk describes reinforcement learning without supervised fine-tuning or distillation in this post-training run, using about 1,200 curated long-horizon tasks over 18 days on 144 H200 GPUs. That is a statement about RadixArk’s training stage, not a claim that the base model had no earlier supervised training.
The card lists 27.8 billion total parameters, including the vision encoder and multi-token-prediction head. It says the architecture, tokenizer and chat template are retained from the base model. This gives existing Qwen3.8 deployments a concrete candidate to evaluate without moving to a much larger mixture-of-experts model, but the vendor’s drop-in description is not a compatibility test of a reader’s runtime.
Training sessions reportedly reached 1 million tokens through harness summarization. That does not make this a million-token inference model: its published context is 262,144 tokens. Likewise, a text-and-image input interface is not evidence of audio, video or image generation support.
What the reported coding improvement does and does not establish
RadixArk reports a Terminal-Bench 4.1 pass@1 increase from 5.6% for the base model to 12.0% for DaisyRose. The launch notes say DaisyRose was evaluated over eight repetitions per task using an internal mini-swe-agent harness with summarization. Those figures are vendor-reported and have not been independently reproduced by AccessAllGPT.
The model card adds an important qualification: some comparison entries in the Terminal-Bench 4.1 row actually use Terminal-Bench 4.0 scores where a 4.1 entry was unavailable. Its baseline sources also involve different harnesses. Do not read that row as a uniform, controlled head-to-head ranking. The card further says DaisyRose scores use the higher of BF16 and FP8 runs where both were evaluated.
For a team already serving the base model, the useful question is whether the new checkpoint completes more of its actual tasks at an acceptable cost. Keep tools and permissions unchanged, replay non-sensitive repository tasks, require tests to pass, and record retries and elapsed time. Better results in a vendor’s agent harness do not justify granting an unfamiliar model broader credentials.
Choose the API or self-hosting path deliberately
The OpenRouter provider record lists BF16 serving, tools, tool choice, structured outputs and optional reasoning, with reasoning enabled by default in the model registry. These are advertised capabilities, not features we exercised. The listed 235,929-token maximum completion is also a provider limit, not a promise that the full context can simultaneously be used for input and output.
For local serving, RadixArk recommends SGLang with DSpark speculative decoding. Its model card puts weight storage at roughly 52 GB for BF16, 29 GB for FP8 and 23 GB for NVFP4, and says NVFP4 requires Blackwell hardware. Those are vendor weight-footprint estimates, not total runtime memory requirements: context caches and serving overhead still need room. We have not verified a particular GPU configuration.
The paid API is the simpler first evaluation path if the data terms fit the workload. Self-hosting becomes a separate operational decision about memory, throughput, isolation and support. Low input pricing alone does not establish low completed-task cost: reasoning output, retries and unsuccessful work still count. Cache pricing is not a blanket discount on every request, and listed rates can change.
A first evaluation plan
Suggested checks for readers, not an experiment AccessAllGPT performed.
Entries stay in this browser tab and are not submitted to AccessAllGPT. Blank responses are copied as [Unresolved].
Confirm radixark/daisyrose-27b and the current provider, rates, quota and data terms in your account.
Replay a small non-sensitive task set against your current model with identical tools, tests and permissions.
Measure billed input, reasoning/output usage, retries and time per accepted result rather than relying on token price alone.
For self-hosting, pin reviewed weights and runtime versions, verify memory under your context length, and test parsers and tool calls.
Primary sources
Browse the publication-wide evidence index →
- DaisyRose-27B: Pushing the Frontier of Post-Training EfficiencyRadixArk · Reviewed: November 18 date, launch and checkpoint links, task design, training infrastructure and evaluation footnote · Retrieved · Supports: Original announcement date and vendor-reported training setup and coding results. Does not independently validate those results.
- DaisyRose-27B model cardRadixArk on Hugging Face · Reviewed: General information, evaluations and comparison footnotes, serving, hardware, model properties and license · Retrieved · Supports: Parameter count, context, modalities, base model, license declaration, linked variants, hardware guidance and benchmark-comparability caveats.
- DaisyRose-27B repository inventoryHugging Face · Reviewed: Public/gated flags, cardData and siblings · Retrieved · Supports: Directly observed public ungated BF16 repository and safetensors shard filenames; not downloaded or integrity-validated weights.
- DaisyRose-27B licenseRadixArk on Hugging Face · Reviewed: License file retrieved through the raw main path · Retrieved · Supports: Published Apache 2.0 license text for this repository.
- DaisyRose 27B on OpenRouterOpenRouter · Reviewed: Overview, release label, modalities and pricing · Retrieved · Supports: Public managed-model discovery page, not proof of successful inference.
- OpenRouter Models APIOpenRouter · Reviewed: radixark/daisyrose-27b record: ID, created, context, pricing, reasoning and modalities · Retrieved · Supports: Registry creation timestamp, exact model ID, per-token price snapshot, context and default reasoning.
- Exact DaisyRose provider endpointOpenRouter · Reviewed: DeepInfra endpoint, quantization, prices, completion limit and supported parameters · Retrieved · Supports: Observed provider listing and advertised API capabilities and limits; no inference call was made.
Limitations
No hands-on inference, downloaded-weight validation, serving experiment, GPU memory measurement or independent benchmark reproduction. The managed route is paid, not a free-model test candidate. Provider availability, limits and pricing are snapshots retrieved on November 21. Other linked checkpoint variants were not individually verified. Cross-model benchmark rows mix versions and harnesses; no biological applications were investigated.
Disclosures
This is public-source reporting, not a vendor-sponsored test or a claim of private access. No vendor reviewed this briefing. AccessAllGPT Research is operated by NeuralArc; publication-wide relationships appear on the disclosures page.



