
DaisyRose 27B vs Gemma 4 31B: Two Dense Models, Two Different Jobs
- 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 Gemma 4 31B are the closest pair in this family of comparisons: both dense, both between 27 and 31 billion parameters, both multimodal with a 256K-token context, both Apache 2.0, both shipping a speculative-decoding drafter alongside the weights. They are also, on the evidence their vendors chose to publish, aimed at different buyers. Gemma 4 31B is Google DeepMind's generalist — a 30.7B dense model with a configurable thinking mode, native function calling, more than 140 languages and a fourteen-row capability table covering everything from AIME to document parsing. DaisyRose 27B is RadixArk's specialist, published on 18 November 2026, post-trained with reinforcement learning alone on long-horizon agent tasks, and proved on a five-row agentic table. The models are near-identical in shape; the proof each vendor offers is nearly disjoint.
Only one benchmark in either table carries both names: GPQA Diamond, where DaisyRose 27B reports 90.7 and Gemma 4 31B reports 84.3. Six and a half points on a graduate-level science exam from a model with 2.9 billion fewer parameters is the headline this page has, and the rest of it is about why that headline is thinner than it looks.
Two architectures that look alike
Strip the branding and the spec sheets overlap more than any other matchup in this series.
• Parameters — DaisyRose 27B 27.8B dense, of which 0.46B is a vision encoder and 0.42B a multi-token-prediction head, against Gemma 4 31B at 30.7B dense with an approximately 550M-parameter vision encoder. No expert routing on either side, so the active-parameter count equals the total on both.
• Context — 262,144 tokens for DaisyRose 27B; 256K for Gemma 4 31B. In practice the same ceiling.
• Modalities — text and image in, text out, on both. Audio is not available at either size; on the Gemma side it is confined to the E2B, E4B and 12B members of the family.
• Licence — Apache 2.0 on both, stated on both model cards.
• Long-context technique — Gemma 4 31B gets there with 60 layers, a 1,024-token local sliding window interleaved with full global attention, unified keys and values in the global layers and proportional RoPE. DaisyRose 27B inherits Qwen3.8-27B's architecture untouched, which is the same hybrid-attention idea arrived at from a different design lineage.
The two places they genuinely differ are the ones that decide what the model is for. Gemma 4 31B ships a thinking mode you can configure, native function calling and explicit support for the system role. DaisyRose 27B ships a chat template identical to its base model's and a training run that spent its entire budget on terminal-and-tool behaviour rather than on breadth.
The single row they share
GPQA Diamond, 198 graduate-level science questions, is the only benchmark both cards publish, and the configurations behind the two numbers are not the same measurement.
DaisyRose 27B's 90.7 comes from eight runs on 198 questions using the simple-evals template, with answer order reshuffled every time, reported as the better of its BF16 and FP8 runs. Gemma 4 31B's 84.3 is a single figure on Google's model card, with no stated run count and no stated harness in the table itself.
So the gap is real in the sense that both vendors published a number under the same benchmark name, and soft in the sense that one side has eight runs and a reshuffle protocol while the other's methodology is a footnote. Six and a half points is roughly thirty-three questions out of 198. That is larger than sampling noise on a 198-question set, but a harness difference of that size is not unheard of, and neither vendor has run the other's evaluation.
What makes the row more interesting than a single score is where it sits relative to each model's base. DaisyRose 27B improves GPQA Diamond from 89.2 to 90.7 over Qwen3.8-27B — a gain of 1.5 points from eighteen days of reinforcement learning. Gemma 4 31B scores 84.3, which is 2 points above its own 26B A4B sibling's 82.3. On this benchmark both models are close to where their pretraining put them, and the post-training moved neither much. If you are choosing between them on graduate science, the six-point gap is mostly a base-model gap rather than a training-recipe gap.

The tables that only one side filled in
Here the comparison stops being a comparison and becomes a choice of evidence. Google published a general capability table and RadixArk published an agentic one, and neither overlaps with the other.
Gemma 4 31B's card carries MMLU Pro 85.2, AIME 2026 without tools 89.2, LiveCodeBench v6 80.0, Codeforces ELO 2150, Tau2 76.9, HLE without tools 19.5 and with search 26.5, BigBench Extra Hard 74.4, MMMLU 88.4, MMMU Pro 76.9, MATH-Vision 85.6, MedXPertQA MM 61.3, OmniDocBench 1.5 at 0.131, and MRCR v2 at eight needles and 128K context 66.4. There is no Terminal Bench row, no SWE-bench row and no IFBench row.
DaisyRose 27B's card carries Terminal Bench 4.1 at 12.0 against a base of 5.6, Terminal Bench 2.1 at 75.5, DeepSWE v1.1 at 46.2, SWE Bench Pro v2 at 94.7 and IFBench at 82.4, plus HLE text-only at 34.3 and GPQA Diamond at 90.7. There is no MMLU Pro, no AIME, no Codeforces and no vision benchmark row at all — despite the model carrying a vision encoder.
Two asymmetries in that pairing deserve naming. The first is HLE: RadixArk reports 34.3 without tools, Google reports 19.5 without tools and 26.5 with search. Even giving Gemma 4 31B its tool-assisted configuration, DaisyRose 27B's number is higher by nearly eight points — but the graders differ, since RadixArk used GPT-6 Luna on five runs per problem while Google does not state its grading set-up. The second is vision: DaisyRose 27B ships a 0.46B vision encoder and publishes no vision benchmark at all, so its multimodal capability is unmeasured in public. Gemma 4 31B publishes MMMU Pro, MATH-Vision, OmniDocBench and MedXPertQA. If you need image understanding, one of these models has evidence and the other has an encoder.
Memory, formats and what each download costs
Both models are 52 to 63 GB at full precision, which puts them in the same hardware class and makes the format list the useful part of the comparison.
• DaisyRose 27B — BF16 about 52 GB, FP8 about 29 GB, NVFP4 about 23 GB and Blackwell-only, plus RadixArk/Qwen3.8-27B-DSpark as a speculative-decoding drafter for SGLang. The reference configuration is four data-parallel replicas across a four-GPU node at tensor parallelism 1, a 262,144-token context and an FP8 KV cache. RadixArk lists single-GPU deployment on H200, B200, GB300, RTX PRO 6000, RTX 5090 and DGX Spark.
• Gemma 4 31B — bfloat16 in two safetensors shards of roughly 49.8 GB and 12.8 GB, about 62.5 GB together. A QAT Q4_0 GGUF at 17.7 GB plus a 1.2 GB vision projector, roughly 18.9 GB before the KV cache, which puts the dense top of the Gemma line on a single 24 GB consumer card. A QAT w4a16 compressed-tensors build at 23.3 GB for vLLM, an unquantised Q4_0 checkpoint carrying the QAT weights without packing, and a 0.94 GB multi-token-prediction drafter.
The GGUF line is the concrete difference. Google ships a 4-bit build of its own making that fits a 24 GB card; RadixArk ships three precisions and expects you to serve them through SGLang. For a team with a workstation and no serving stack, that is not a close call — one download runs, the other needs a deployment.
One detail that bites: Gemma's card requires a speculative-decoding assistant to match the target checkpoint's precision, so the 0.94 GB drafter pairs with the QAT 31B target rather than the bfloat16 release. RadixArk's drafter is named for Qwen3.8-27B and pairs with DaisyRose 27B's own FP8 build.

Cost, and the one that has a price
Gemma 4 31B is on OrcaRouter today as google/gemma-4-31b-it at $0.13 per million input tokens and $0.38 per million output, with configurable thinking, native function calling and a 256K context. The -it suffix is part of the string a client passes.
DaisyRose 27B has no vendor-hosted price. RadixArk's site describes training and serving for other labs and publishes no DaisyRose endpoint; the model card points only at self-hosting through SGLang, vLLM and Docker. Third-party listings quote $0.15 per million input and $1.88 per million output, which is one host's opening position rather than a vendor rate.
Put those side by side and the hosted comparison runs the wrong way from the benchmark table: Gemma 4 31B at thirteen cents per million input against a third-party host's fifteen, but thirty-eight cents against a dollar-eighty-eight on output. On a generation-heavy workload that is a five-fold difference in favour of the bigger model, and it exists for a structural reason — Gemma 4 31B has an established hosted tier with several providers competing on it, while DaisyRose 27B has one host and three days of history.

Which job each model was built for
The honest summary is that these models are not competing for the same purchase, and the benchmark tables prove it by refusing to overlap.
Gemma 4 31B is the model you want if you need breadth: a thinking mode you can turn down, function calling that works with existing tool schemas, 140-plus languages, a published vision evaluation, a quantised build that runs on a workstation, and a hosted price of thirteen cents per million input tokens. It is the safe choice for a general product, and its cost of being wrong is a few dollars of API spend.
DaisyRose 27B is the model you want if you are building a terminal agent and you can host it. Its Terminal Bench 4.1 figure more than doubles its own base model's, its SWE Bench Pro v2 sits at 94.7, and its IFBench at 82.4 is the highest on RadixArk's comparison card. Every one of those numbers is vendor-reported, none has been independently reproduced, and the model is three days old. But the shape of the claim is specific enough to test: run your own agent traces against Qwen3.8-27B — available as qwen/qwen3.8-27b at $0.33 per million input and $2.40 per million output, the same 262,144-token context and vision tower — and see whether the failure mode RadixArk trained away is the one you actually hit.
That is also the cheapest way to decide between these two. Gemma 4 31B and Qwen3.8-27B sit behind one OrcaRouter key with no second contract and no code change, at 0% markup over each provider's list price, so a price change on either side reaches you the day it lands. If Gemma 4 31B handles your agent loop, you are done at thirteen cents. If it does not, and the base model fails in the same place, then an eighteen-day RL post-train on an H200 is the next thing worth paying for — and by then you will know what you are buying it to fix.
