
DaisyRose 27B vs Granite 4.2 30B: 46 Points, Two Harnesses
- OrcaNEWOrcaCyber: Hy4 Flash Abliterated / Uncensored2026-11-18$0.07 / $0.28 per 1M tokens · 184 tok/s
- tencentNEWTencent: Hy4 Flash2026-11-16$0.02 / $0.07 per 1M tokens · 494 tok/s
- AlibabaNEWQwen: Qwen4 Max2026-11-16$1.50 / $4.50 per 1M tokens · 385 tok/s
- GoogleNEWGoogle: Gemini 4 Carbon2026-11-10$2.00 / $10.00 per 1M tokens · 94 tok/s
- AnthropicNEWAnthropic: Claude Fable 5.22026-11-09$8.00 / $40.00 per 1M tokens · 290 tok/s
- OrcaOrcaCyber: GLM 5.3 Abliterated / Uncensored2026-11-04$3.00 / $5.00 per 1M tokens · 99 tok/s
- OrcaOrcaCyber: Hy4 Abliterated / Uncensored2026-11-04$0.80 / $1.60 per 1M tokens · 132 tok/s
- OrcaOrcaCyber: Kimi K3.1 Pro Abliterated / Uncensored2026-11-04$2.00 / $5.00 per 1M tokens · 86 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 · 104 tok/s
- OrcaOrcaCyber: GLM 5.3 Flash Abliterated / Uncensored2026-10-27$0.80 / $1.60 per 1M tokens · 66 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 · 41 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 · 215 tok/s
- minimaxMiniMax: M3.1 Flash2026-10-1929Intelligence
- OrcaOrca: OrcaCyber Ultra 1.02026-10-17$5.00 / $10.00 per 1M tokens · 109 tok/s
- deepseekDeepSeek: DeepSeek V4.1 Flash 10172026-10-1733Intelligence
- DeepSeekDeepSeek: DeepSeek V4.1 Flash 09102026-10-17$0.15 / $0.60 per 1M tokens · 160 tok/s
- DeepSeekDeepSeek: DeepSeek V4.1 Flash2026-10-1740Intelligence
DaisyRose 27B and Granite 4.2 30B both report a score on Terminal Bench 2.1. DaisyRose 27B, RadixArk's reinforcement-learning-only post-train of Qwen3.8-27B published on 18 November 2026, scores 75.5 against its base model's 73.0. Granite 4.2 30B, IBM's dense 30-billion-parameter model released on 25 August 2026, scores 29.24. That is a 46-point gap on the same benchmark name, between two dense models of nearly identical size, and the natural conclusion — that DaisyRose 27B is two and a half times better at agentic terminal work — is almost certainly false. The two numbers were not produced the same way, and this page is largely about how a 46-point gap can be smaller than it looks.
The two models have more in common than that gap suggests: both dense, both around 30 billion parameters, both Apache 2.0, both open-weight releases with no vendor-hosted API price to quote, both trained specifically for agentic tool use inside sandboxes. They are the closest-matched pair in this series by architecture and by intent, which is exactly why the numbers on their cards are worth reading carefully rather than subtracting.
Anatomy of a 46-point gap
Start with what each vendor says about how the number was made, because both cards are unusually explicit and they describe different measurements.
• DaisyRose 27B — five runs per task on the Terminus 2 harness. The comparison card lists Terminus 2 explicitly and states the run count. Sampling was temperature 1.0, top_p 0.95, top_k 20, reasoning effort xhigh, and the published score is the better of BF16 and FP8 runs.
• Granite 4.2 30B — 29.24 on Terminal-Bench 2.1, reported as a single table row on the model card. The card documents the family's agentic training in detail — a second supervised fine-tuning pass over up-sampled agentic software-engineering data, then agentic reinforcement learning inside a sandbox doing SWE-bench-style tasks, terminal operations and web search — but the Terminal-Bench row itself carries no stated run count, no harness name and no sampling configuration.
That asymmetry is the whole story of the 46 points. One vendor published a heavily instrumented score at maximum reasoning effort over five runs; the other published a table cell. A Terminal-Bench result at reasoning effort xhigh can move by tens of points against the same model at default settings, because the benchmark is a sequence of shell commands where a model that thinks longer gets more attempts at recovering from its own mistakes. Comparing a best-effort, high-effort, five-run figure to an unqualified single number is not a comparison at all.
There is a second asymmetry, subtler and more interesting. DaisyRose 27B's own base model, Qwen3.8-27B, scores 73.0 on the same row of the same card. So RadixArk's post-training improved Terminal Bench 2.1 by 2.5 points, while the 46-point gap to Granite is 44 points of base-model difference and 2 points of training. Everything that separates these two vendors on this benchmark is in the starting point, not in the reinforcement learning.
Where the tables do meet
Two more rows are shared, and both are more comparable than Terminal-Bench.
• GPQA Diamond — DaisyRose 27B 90.7 across eight runs on 198 questions with reshuffled answer order; Granite 4.2 30B 66.41. A 24-point gap that no harness difference plausibly explains. DaisyRose's base scores 89.2, so the post-train added 1.5 points to a base model that was already far ahead.
• IFBench — DaisyRose 27B 82.4 at four runs on 300 prompts with the official allenai scorer, strict prompt-level accuracy excluding thinking; Granite 4.2 30B 77.17, labelled "prompt" on the card. A 5-point gap, and the closest thing to a genuine tie on either card.
The GPQA figure is the one that should change how a reader thinks about this matchup. Granite 4.2 30B is a model IBM built for enterprise agentic work — twelve tested languages, three reasoning modes, tool calling, a 128K native context extensible to 512K — and it scores 66.41 on graduate-level science. DaisyRose 27B scores 90.7. Whatever the harness caveats on Terminal-Bench, graduate science questions do not have a harness that swings 24 points, and the honest reading is that RadixArk's base model, Qwen3.8-27B at 89.2, is simply a stronger reasoning model than the base IBM started from.

Two very different ideas of agentic training
Both models were trained to work inside a sandbox, and the two descriptions of that training are the most revealing part of either card.
RadixArk's approach was to change as little as possible. DaisyRose 27B keeps Qwen3.8-27B's architecture, tokenizer, chat template, vision tower and 262,144-token context untouched, and applies reinforcement learning to roughly 1,200 curated long-horizon tasks built with Eigent AI — tasks designed to chain several problems, resist reward hacking, and run in sessions of up to one million tokens. Eighteen days, 144 H200s, 63,000 GPU-hours, about 944,000 sandboxes. No supervised fine-tuning, no distillation.
IBM's approach was to change the behaviour in stages. Granite 4.2 30B's family received a second supervised fine-tuning pass over up-sampled agentic software-engineering data, then agentic reinforcement learning that ran the model inside a sandbox on SWE-bench-style tasks, terminal operations and web search. The card is explicit that the 30B is the family's strongest performer on reasoning-intensive and agentic work. It also carries a full general-capability table, twelve tested languages, three thinking modes and built-in chain-of-thought.
Neither vendor's training description predicts the other's benchmark profile, and that is what the tables show: DaisyRose 27B leads decisively on the reasoning-heavy rows and publishes nothing on multilingual or vision; Granite 4.2 30B publishes across the enterprise suite — MMLU-Pro 77.60, AIME25 89.17, LiveCodeBench v6 75.77, Arena-Hard-V2 67.93, BFCL v4 61.39, τ³-bench 62.00, SWE-bench Verified 57 — and scores 66.41 where the two overlap on science.
Context, memory and the deployment story
Granite's 512K extension is a real differentiator that no benchmark row captures, and it is worth being precise about what the card claims: 128K native, "long-context extension to 512K", with the model table listing a sequence length of 131,072. Extension is not the same as native, and the RULER figures on the card show the cost — 89.96 at 64K against 81.38 at 128K on the same model. DaisyRose 27B takes the opposite position: a flat 262,144 tokens with no long-context variant, inherited from Qwen3.8-27B, and RadixArk notes that its own long thinking runs reach tens of thousands of tokens, which is where much of that window goes in practice.
• DaisyRose 27B weights — BF16 about 52 GB, FP8 about 29 GB, NVFP4 about 23 GB and Blackwell-only, plus the Qwen3.8-27B-DSpark speculative-decoding drafter. RadixArk recommends SGLang with DSpark and publishes a four-GPU reference configuration with an FP8 KV cache.
• Granite 4.2 30B weights — about 29B stored in bfloat16, Apache 2.0, with FP8, NVFP4/MXFP4 and GGUF quantisations shipped alongside the BF16 release. At 30B dense it lands near the same memory class as DaisyRose's BF16 build, so neither has a footprint advantage worth arguing about.
The deployment difference is one of support, not of size. Granite 4.2 30B has been downloadable and runnable for three months, has quantisations published by IBM for several runtimes, and has an ecosystem of people who have already hit its failure modes. DaisyRose 27B is three days old, has three precisions, one recommended serving path and no published throughput figure for any hardware configuration.

Neither one is on a hosted endpoint here
This is the least satisfying section of the comparison and it is worth stating plainly: neither DaisyRose 27B nor Granite 4.2 30B appears on OrcaRouter's catalogue at the time of writing. RadixArk publishes no hosted DaisyRose API and points at SGLang, vLLM and Docker; IBM's Granite 4.2 30B is an open-weight release whose practical path is self-hosting, with no vendor rate to quote. Third-party listings for DaisyRose quote $0.15 per million input tokens and $1.88 per million output, which is a host's opening position rather than a vendor price and should be read as one.
What that means for a reader trying to decide is that this comparison cannot be settled by switching a base URL. Both models want a GPU, and the honest advice is to run the cheaper experiment first: if your workload fits Granite 4.2 30B's 128K native context and its published capability table, the model has three months of other people's mistakes behind it and you can find out in a day whether it fails on your tasks.
What would make this comparison real
One run on one harness would settle most of this page. Terminal-Bench is public, the tasks are fixed, and both models load in llama.cpp-adjacent runtimes at comparable precision. If someone published DaisyRose 27B and Granite 4.2 30B on the same harness at the same sampling settings, the 46-point gap would collapse into whatever it actually is — probably somewhere between the two vendors' figures, since DaisyRose's is a best-of-two at xhigh effort and Granite's is unqualified.
Until then, the defensible reading is narrower and still useful: on graduate-level science, where harnesses matter least, RadixArk's model is far ahead of IBM's, by a margin that neither run count nor answer-order shuffling can account for. On instruction following the two are close. On terminal agent work nobody can say, and the published 46-point gap measures two evaluation setups at least as much as it measures two models.
For a team choosing today, the decision is less about the benchmark table than about maturity. Granite 4.2 30B is a finished, documented, three-month-old enterprise model with quantisations and a multilingual table. DaisyRose 27B is an eighteen-day reinforcement-learning experiment with a striking set of vendor numbers, one shared row that supports them, and no independent reproduction of any kind. If the workload is production and the deadline is now, that difference decides it. If the workload is an agent you control and you can afford to benchmark, DaisyRose 27B's GPQA and IFBench rows are reason enough to put it on the same harness as the model you are running today and see what happens.
