
DaisyRose 27B: 18 Days of RL and No Supervised Fine-Tuning
- OrcaNEWOrcaCyber: Hy4 Flash Abliterated / Uncensored2026-11-18$0.07 / $0.28 per 1M tokens · 187 tok/s
- openaiNEWOpenAI: GPT-6.2 Astra2026-11-17$8.00 / $40.00 per 1M tokens · 87 tok/s
- openaiNEWOpenAI: GPT-6.2 Sol2026-11-17$2.00 / $10.00 per 1M tokens · 152 tok/s
- tencentNEWTencent: Hy4 Flash2026-11-16$0.02 / $0.07 per 1M tokens · 194 tok/s
- AlibabaNEWQwen: Qwen4 Max2026-11-16$1.50 / $4.50 per 1M tokens · 86 tok/s
- GoogleNEWGoogle: Gemini 4 Carbon2026-11-10$2.00 / $10.00 per 1M tokens · 705 tok/s
- AnthropicNEWAnthropic: Claude Fable 5.22026-11-09$8.00 / $40.00 per 1M tokens · 246 tok/s
- OrcaOrcaCyber: GLM 5.3 Abliterated / Uncensored2026-11-04$3.00 / $5.00 per 1M tokens · 97 tok/s
- OrcaOrcaCyber: Hy4 Abliterated / Uncensored2026-11-04$0.80 / $1.60 per 1M tokens · 138 tok/s
- OrcaOrcaCyber: Kimi K3.1 Pro Abliterated / Uncensored2026-11-04$2.00 / $5.00 per 1M tokens · 85 tok/s
- tencentTencent: Hy42026-11-04$0.83 / $2.50 per 1M tokens · 122 tok/s
- OrcaOrcaCyber: MiniMax M3.1 Flash Abliterated / Uncensored2026-10-29$0.80 / $1.60 per 1M tokens · 102 tok/s
- OrcaOrcaCyber: GLM 5.3 Flash Abliterated / Uncensored2026-10-27$0.80 / $1.60 per 1M tokens · 70 tok/s
- openaiOpenAI: GPT-6.1 Terra2026-10-26$1.00 / $5.00 per 1M tokens · 35 tok/s
- OrcaOrcaCyber: Qwen3.8 Flash Next Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 89 tok/s
- OrcaOrcaCyber: MiMo V2.6 Flash Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 38 tok/s
- MoonshotMoonshotAI: Kimi K3.1 Pro2026-10-2447Intelligence
- OrcaOrcaCyber: DeepSeek V4.1 Flash Abliterated / Uncensored2026-10-23$0.80 / $1.60 per 1M tokens · 231 tok/s
- minimaxMiniMax: M3.1 Flash2026-10-1929Intelligence
- OrcaOrca: OrcaCyber Ultra 1.02026-10-17$5.00 / $10.00 per 1M tokens · 101 tok/s
DaisyRose 27B from RadixArk is a post-train of Qwen3.8-27B built with reinforcement learning and nothing else — no supervised fine-tuning pass, no distillation from a larger teacher — and the two benchmark rows that best describe what that buys sit four lines apart on its own model card. On Terminal Bench 4.1 the model goes from 5.6 to 12.0 against its base, better than doubles. On DeepSWE v1.1 it goes from 42.2 to 46.2, a gain of four points. One row says the recipe taught the model something it could not do at all; the other says most of what it already could do, it does about as well as before. RadixArk published the weights and the recipe on 18 November 2026 under Apache 2.0, and the honest reading of the release is that both rows are true at once.
This is a launch, and a recent one — the blog post and the Hugging Face card are three days old at the time of writing. What follows is what RadixArk claims, which parts of it are checkable today, and where the numbers come from.
What was actually trained
The base is Qwen3.8-27B, unchanged in architecture: same tokenizer, same chat template, same vision tower, same 262,144-token context. RadixArk describes DaisyRose 27B as a drop-in replacement for that base, which is a claim about compatibility rather than quality — existing prompts, image inputs and serving configs carry over without modification.
• Total parameters — 27.8B, of which 0.46B is the vision encoder and 0.42B is a multi-token-prediction head.
• Training signal — reinforcement only. RadixArk states the run used no supervised fine-tuning and no distillation, which is unusual at this size; most open post-trains at 27B start from an SFT pass and use RL to sharpen it.
• Environment — roughly 1,200 curated long-horizon agent tasks, built with Eigent AI, designed to chain several problems together and to be resistant to reward hacking. Sessions ran up to one million tokens, which is the number that explains the rest of the setup: a task that long needs a sandbox that stays alive for hours.
• Compute — 18 days on 144 H200s, split 128 for inference and 16 for the trainer, running fully asynchronously, for about 63,000 GPU-hours. The trainer is the Miles framework.
• Sandboxes — about 944,000 of them over the run, peaking at 1,566 concurrent, with over 99.95% creation success in late October. RadixArk puts that infrastructure at roughly 80% below what it says popular cloud sandboxes charge.
The 128/16 split is the part worth pausing on. In an asynchronous RL loop the inference fleet generates rollouts for long agent sessions while the trainer consumes finished trajectories, so the ratio is a statement about how long each rollout takes relative to a gradient step. It is also why RadixArk's efficiency pitch is framed in wall-clock rather than in FLOPs: the run finished in eighteen days, and the number of GPUs is unremarkable next to that.
The benchmark table, read with its footnotes
RadixArk reports seven benchmarks against the Qwen3.8-27B baseline, and states per benchmark how many runs, which harness, and which sampling settings. That footnoting is more than most launch cards do, and it is the reason the numbers below can be compared to each other at all. They are still vendor-run and nobody outside RadixArk has reproduced them.
• Terminal Bench 4.1 — 12.0 against 5.6 for the base. Eight runs per task on an internal mini-swe-agent harness with summarisation enabled. Models with no 4.1 entry in the cross-model comparison have their 4.0 score substituted.
• Terminal Bench 2.1 — 75.5 against 73.0. Five runs per task, Terminus 2 harness. The gain here is under three points, which is the opposite impression from the 4.1 row: the 2.1 suite is the one the base model was already good at.
• DeepSWE v1.1 — 46.2 against 42.2. Eight runs, mini-swe-agent.
• SWE Bench Pro v2 — 94.7 against 91.3. Five runs, mini-swe-agent, all 642 tasks. Several of the compared models report only the Hard subset, so this row is not apples to apples across columns.
• HLE, text only — 34.3 against 33.9. Five runs per problem, no tools, graded by GPT-6 Luna.
• IFBench — 82.4 against 79.5. Four runs on 300 prompts with the official allenai scorer, strict prompt-level accuracy excluding thinking.
• GPQA Diamond — 90.7 against 89.2. Eight runs on 198 questions, simple-evals template, answer order reshuffled each time.

Two properties of that table matter more than any single row. The first is that every DaisyRose score is the better of its BF16 and FP8 runs, which RadixArk says differ by under one point — a small margin, but it means the published figure is a best-of-two rather than a mean. The second is sampling: temperature 1.0, top_p 0.95, top_k 20, reasoning effort xhigh. High effort and high temperature are a configuration chosen to maximise agentic scores, and a reader running at default settings should expect something lower than the table.
Where it is not the leader
The model card compares DaisyRose 27B against seven other systems, and the result is not a clean sweep — which is the most useful thing about the table, because the losses are informative.
• Terminal Bench 4.1 — MiMo V2.6 Flash scores 16.1, Inkling 2 Small 13.0, GPT-6 Luna 15.2, all above DaisyRose's 12.0. Muse Glimmer-30B scores 0.5 and Qwen3.6-35B-A3B scores 0.0 on the same harness.
• Terminal Bench 2.1 — MiMo V2.6 Flash 87.6 and Inkling 2 Small 81.3 both lead; Muse Spark Open sits at 76.4.
• DeepSWE v1.1 — Inkling 2 Small 72.1, MiMo V2.6 Flash 67.9 and GPT-6 Luna 66.6 against DaisyRose's 46.2. This is the widest loss on the card, and it lands on the benchmark that most closely resembles the training distribution.
• HLE text-only — Inkling 2 Small 44.1 and Muse Spark Open 43.9, against 34.3.
• GPQA Diamond — Inkling 2 Small 93.0, Beam 1.0 90.6, GPT-6 Luna 90.5, DaisyRose 90.7. A near-tie at the top.
• IFBench — DaisyRose 82.4 is the highest figure on the card; no competitor row has a number.
What that pattern says is that the 27.8B dense model wins on instruction following and holds its ground on graduate-level science, and loses to much larger sparse models on the deep software-engineering tasks. MiMo V2.6 Flash activates 15B of 309B and Inkling 2 Small activates 17B of 594B, so the comparison is a 27B dense against models with twenty times the stored parameters, routed. Losing to those is not a failure of the recipe; it is a statement about how much of the gap a 27B can close.

The 37% MFU claim and what it is measuring
RadixArk's blog leads with training efficiency rather than with the model, which is a fair read of what the company sells — it is an infrastructure outfit that has served SGLang deployments for other labs — but it makes the efficiency numbers harder to price than a benchmark row.
The claim is 37% model FLOPs utilisation for its internal backend against 16% for a Megatron-based open-source trainer, described as 57% lower training cost and a 2x end-to-end speedup. There is no published configuration for the Megatron baseline: no parallelism plan, no sequence length, no batch size. A 16% MFU figure is plausible for a naive dense training loop and implausible for a well-tuned one, so the honest status of "57% lower cost" is that it is a comparison against an unnamed configuration of an unspecified trainer, measured by the party that built the alternative.
The 63,000 GPU-hours, by contrast, is a concrete figure and it can be sanity-checked against the wall clock: 144 GPUs over 18 days is 62,208 GPU-hours, which matches. So the run cost what RadixArk says it cost in machine time. What the efficiency claim adds on top is a statement about what that machine time would have cost someone else, and that part is unverified.
What you can download, and what it takes to serve
RadixArk published three precisions of the weights, all Apache 2.0, all on Hugging Face, plus a speculative-decoding drafter.
• BF16 — about 52 GB. Full precision, and the reference for reproducing the card's numbers.
• FP8 — about 29 GB. The build RadixArk's own serving example runs, with an FP8 KV cache.
• NVFP4 — about 23 GB, and Blackwell-only. On an RTX 5090 or a DGX Spark this is the build that makes a 27B multimodal model fit comfortably with room for a long context.
• Drafter — RadixArk/Qwen3.8-27B-DSpark, a speculative-decoding draft model for SGLang, which is the serving path the card recommends. The reference configuration runs four data-parallel replicas across a four-GPU node at tensor parallelism 1, a 262,144-token context and an FP8 KV cache.
Hardware-wise RadixArk lists single-GPU deployment on H200, B200 and GB300, plus RTX PRO 6000, RTX 5090 and DGX Spark. No throughput or tokens-per-second figure is published anywhere on the card, which is a real gap for anyone sizing a deployment: the one number that would tell you whether DSpark is worth the extra GPU is missing.
Availability, and the pricing question
There is no RadixArk-hosted API price to quote. The company's site describes serving and training for other labs and publishes no DaisyRose endpoint, and the model card points only at self-hosting through SGLang, vLLM and Docker. Third-party listings do exist: one we saw at publication quotes $0.15 per million input tokens and $1.88 per million output tokens, which is a third-party host's number rather than a vendor list price and should be treated as one host's opening position rather than the market rate.

DaisyRose 27B itself is not on OrcaRouter's catalogue at the time of writing, and we are not going to imply otherwise. Its base model is: Qwen3.8-27B is served under qwen/qwen3.8-27b, Alibaba's own open-weight build, at $0.33 per million input tokens and $2.40 per million output, with a 262,144-token context and the same vision tower the post-train keeps. That matters for anyone evaluating this release, because the cheapest way to test whether a 12-point Terminal Bench gain is worth standing up an H200 is to run the base through an API first, on the same prompts, and compare. It also gives a fallback path: if a self-hosted DaisyRose deployment falls over mid-task, the base model sits behind the same key, and an automatic failover rule can hand the request to it rather than failing the call.
What would settle it
Nothing in this release has been independently run. There is no Artificial Analysis entry for DaisyRose 27B, no third-party harness on the Terminal Bench leaderboard, and the cross-model rows in RadixArk's own comparison draw on other parties' published figures rather than on runs performed side by side. The card is unusually explicit about all of this, which is to its credit, but explicit provenance does not make a number independent.
Two things are worth watching. The first is a reproduction of Terminal Bench 4.1 on the public harness: 5.6 to 12.0 is the claim the whole release rests on, and it is the kind of jump that a harness difference can also produce. The second is DeepSWE, where the model sits 26 points behind Inkling 2 Small despite the training data being agentic software engineering. If the 4.1 gain reproduces and the DeepSWE gap does not close, the story is that RL-only post-training teaches a model to use a terminal without teaching it to reason about a codebase — and that is a much more specific, and more useful, finding than any single headline score.
Compared in this article1
Detected from this article · Benchmarks: Artificial Analysis · updated daily
