
DaisyRose 27B vs ZWZ2 27B: Averages vs Frontiers
- 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
Put DaisyRose-27B and ZWZ2-27B next to each other and you get two genuine reinforcement-learning post-trains of Qwen checkpoints, released three weeks apart in the autumn of 2026, reported in two completely different styles. DaisyRose-27B, from RadixArk on 18 November 2026, is seven benchmarks against Qwen3.8-27B, with the largest relative gain sitting on the hardest one. ZWZ2-27B, from inclusionAI, is twenty-five benchmarks across six domains against Qwen3.6-27B, reported as a mean score that rises from 55.52 to 60.51.
One lab published a frontier. The other published an average. Those are not the same kind of claim, and the difference tells you more about what each model is for than any single number in either table.
Different bases, one generation apart
The single most important fact in this matchup is that the two models do not share an ancestor. ZWZ2-27B is built on Qwen3.6-27B; DaisyRose-27B is built on Qwen3.8-27B. That is one base generation apart, and it means the head-to-head that a "vs" page normally performs is not available here at all — not because the numbers are missing, but because beating a previous-generation base says something different from beating the current one.
ZWZ2-27B's own card makes this clear by reporting every gain against Qwen3.6-27B. DaisyRose-27B's card reports against Qwen3.8-27B. A reader who compares the two uplift figures is comparing an improvement over an older base with an improvement over a newer one, and the newer base started higher. No arithmetic fixes that, and no honest article should try.
How ZWZ2 reports itself
ZwZ2-27B's evaluation is organised by domain rather than by benchmark difficulty. Twenty-five benchmarks grouped into six areas, each domain carrying its own before-and-after average:
• Fine-grained perception — 62.09 to 67.57, a gain of 5.48 across eight benchmarks including VStar, HR-Bench 8K and MME-RealWorld.
• Spatial reasoning — 49.72 to 55.98, up 6.26, the largest domain gain.
• General visual reasoning — 47.89 to 53.71, up 5.82, spanning MMStar and BabyVision.
• Embodied spatial — 50.17 to 55.71, up 5.54.
• Long context — 65.87 to 68.77, up 2.90.
• Multimodal agents — 49.73 to 51.81, up 2.08, the smallest gain and the most telling one.
The overall mean moves from 55.52 to 60.51, a rise of 4.99, and inclusionAI states that the tuned model beats the base on every one of the twenty-five benchmarks. That is a strong claim about consistency and a weak claim about any particular capability, because a mean over twenty-five tests is dominated by whichever tests the model was already good at.
How DaisyRose reports itself
RadixArk took the opposite approach: seven benchmarks, each reported individually, with the base model's score beside it. Terminal Bench 4.1 goes from 5.6 to 12.0, Terminal Bench 2.1 from 73.0 to 75.5, DeepSWE v1.1 from 42.2 to 46.2, SWE Bench Pro v2 from 91.3 to 94.7, Humanity's Last Exam from 33.9 to 34.3, IFBench from 79.5 to 82.4, GPQA Diamond from 89.2 to 90.7.
No mean is offered, and that is the point. The headline gain, more than double on Terminal Bench 4.1, is a gain from a very low base — 5.6 is close to floor. The two benchmarks where the base was already strong, HLE and GPQA Diamond, barely move. Read as a set, the table says the tune bought agentic competence and left general reasoning alone, which is exactly what a reward built on long-horizon agent tasks should produce.

What a 25-benchmark mean hides
ZWZ2-27B's weakest absolute rows are the interesting ones, and its own card publishes them: Chartography 12.00, AgentVista 18.18, Perception-Bench 44.93 and ScreenSpot-Pro 70.59. Those sit alongside domain averages in the fifties and sixties, which means the mean is being held up by the benchmarks the model does well and the genuinely hard ones are near floor.
That is normal for a post-train and should not be read as a flaw. It is, however, precisely the information a mean erases. A team that reads "60.51" and expects a model that is 60% good at everything will be badly surprised by Chartography at 12. A team that reads the per-benchmark rows knows immediately which of its workloads this model was not trained for.
The same caveat applies in reverse to DaisyRose-27B. Its Terminal Bench 4.1 score of 12.0 is a large relative gain and a poor absolute result, and the card is honest about both. Neither model's headline number survives contact with its own detail rows.
Architecture and licence
Both inherit their base unchanged. ZWZ2-27B carries the Qwen3.6-27B config as published: hidden size 5,120, 64 layers, 24 query heads and 4 key-value heads at dimension 256, an FFN of 17,408, a repeated three-to-one linear-to-full attention pattern sixteen times over, interleaved M-RoPE with a rope base of ten million, a vocabulary of 248,320, and 262,144 native positions. Its vision tower is 27 layers at width 1,152 projecting to 5,120 with 16-pixel patches and a 2×2 merge. There is no multi-token-prediction layer — the config sets that to zero — and the BF16 weights run to roughly 54 GB across two shards, with transformers 5.13.1 as the documented path.
DaisyRose-27B carries the Qwen3.8-27B config: same hidden size and layer count, Gated DeltaNet linear attention in the same three-to-one pattern, a 0.46B vision encoder, and a 0.42B prediction head that ZWZ2-27B does not have. Context is 262,144 for both. DaisyRose ships in BF16 at about 52 GB, FP8 at 29 GB and NVFP4 at 23 GB with a full SGLang serving recipe; ZwZ2-27B ships BF16 only.
• Licence — Apache-2.0 for both, inherited from their bases.
• Vision — both accept images and return text; neither is text-only.
• Quantisations — three formats for DaisyRose-27B, BF16 only for ZwZ2-27B.
• Serving documentation — a detailed SGLang recipe with speculative decoding for DaisyRose-27B; a transformers quickstart for ZwZ2-27B.
• Independent evaluation — none for either model at the time of writing.

The perception question
ZwZ2-27B's gains concentrate in perception and spatial reasoning: fine-grained perception up 5.48, spatial up 6.26, embodied spatial up 5.54, high-resolution benchmarks like HR-Bench 8K sitting inside that first group. DaisyRose-27B reports no perception benchmark at all. Its vision encoder is inherited and untouched, and nothing in the card suggests the reward shaped it.
That asymmetry is the decision. If your workload is a screenshot, a dense chart, a scanned page or a spatial layout, ZwZ2-27B's card addresses it directly and DaisyRose-27B's does not. If your workload is a terminal, a repository or a long tool-call session, the reverse is true: DaisyRose-27B's card is full of exactly those rows and ZwZ2-27B's weakest domain is multimodal agents at a mean of 51.81.
Neither card is better evidence than the other. They are evidence about different things, produced under different protocols, and the right response is to read each one for the question it was built to answer.
Running them, and mixing them
Neither DaisyRose-27B nor ZwZ2-27B is hosted on OrcaRouter today. Both are weights you serve yourself or reach through third-party platforms, which means both carry the same operational problem: two checkpoints, two deployments, two sets of capacity decisions, and no way to send a request to whichever one is currently available.
That is the case routing was built for. When two models split a workload along a clean line — perception on one side, agentic action on the other — the application ends up classifying requests and calling two different endpoints anyway. OrcaRouter's routing DSL lets that classification live in one place, composing the two models into a single call path, with automatic failover so that a saturated self-hosted deployment falls through to a second endpoint rather than returning an error. For a pair of models this new and this thinly documented, failover is less about uptime than about not betting a production path on weights that have been public for three weeks. OrcaRouter also passes provider list price through at 0% markup for everything we do host, so the hosted leg of a mixed deployment is priced at cost rather than at a multiple.

Which report to trust for your decision
Take ZwZ2-27B if your work is perception-first and you want a model whose card documents it across twenty-five benchmarks — and read the absolute rows, not the mean. Take DaisyRose-27B if your work is agentic and you want a model whose card documents a specific, verifiable behavioural change against the base it came from. Take neither if you need independent verification before committing: both tables are vendor-reported, neither has an Artificial Analysis score, and the one number a procurement team would want — performance on your own workload — is in neither document.
The comparison that will actually matter is the next one. Both labs are on their first model, both built on Qwen checkpoints, and both used reinforcement learning as the whole post-training story. If inclusionAI ships a ZwZ3 on Qwen3.8-27B, or RadixArk ships a second DaisyRose that keeps the agent recipe and adds perception rows, the two reporting styles will finally meet on a shared base and a real head-to-head becomes possible. Until then, the correct summary is that these models were measured for different purposes, and the difference between an average and a frontier is the difference between asking "is it better overall" and asking "is it better at the thing I need".
