
DaisyRose 27B vs Gemma 4 26B A4B: 27.8B Dense vs 3.8B Active
- 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 26B A4B are the same weight of model with completely different cost structures, and the price line makes that concrete faster than any benchmark. Gemma 4 26B A4B — 25.2 billion parameters, of which 3.8 billion are read for each token — is served on OrcaRouter at $0.06 per million input tokens and $0.33 per million output. DaisyRose 27B, RadixArk's reinforcement-learning-only post-train of Qwen3.8-27B published on 18 November 2026, is 27.8 billion dense parameters with every one of them read on every token, ships as 52 GB of BF16 weights, has no vendor-hosted API, and needs a datacentre GPU to run at all. One is a model you rent by the cent. The other is a model you deploy. Everything else in this comparison follows from that.
The benchmark picture is genuinely interesting and genuinely incomplete: the two models share exactly two public rows, GPQA Diamond and Humanity's Last Exam, and on both of them DaisyRose 27B is ahead by a margin much wider than the parameter difference. Whether that margin survives contact with a shared harness is the question this page cannot answer, and says so.
The name is a specification
Gemma 4 26B A4B's label carries its routing policy, and Google's own card states the reading directly: "the 'A' in 26B A4B stands for 'active parameters' in contrast to the total number of parameters the model contains."
• Total parameters — 25.2B. • Active per token — 3.8B. • Expert layout — 8 of 128 experts active plus 1 shared expert, selected per token. The shared expert is always read, which is why the active count is not simply 8/128ths of the total. • Layers — 30, with a 1,024-token sliding window.
DaisyRose 27B has no such structure to describe. It is a dense transformer: 27.8B parameters, 0.46B of them a vision encoder and 0.42B a multi-token-prediction head, with no router between the input and the compute. Every token pays for every weight.
The practical consequence is the one number that separates these models more than any benchmark. A sparse model at 3.8B active runs at something close to the speed of a 4B model while carrying the memory of a 25B one — Google's card says as much outright, describing the MoE variant as running "almost as fast as a 4B-parameter model." A dense 27.8B runs at the speed of a 27.8B, and nobody has published that speed: RadixArk's card carries no tokens-per-second figure for any hardware configuration.
What the RL bought, and what it did not
DaisyRose 27B's training is the part of this comparison with no analogue on the Gemma side. RadixArk used reinforcement learning only — no supervised fine-tuning pass, no distillation — over roughly 1,200 curated long-horizon agent tasks built with Eigent AI, with sessions running up to one million tokens, for eighteen days on 144 H200s and about 63,000 GPU-hours. The aim was not a general capability lift but a specific one: make a 27B better at the loop of read a task, call a tool, read the result, fix the mistake.
That shows up in the rows where the two models can actually be compared, and it does not show up uniformly.
• GPQA Diamond — DaisyRose 27B 90.7 against Gemma 4 26B A4B's 82.3. Both vendor-reported: RadixArk's figure is across eight runs on 198 questions with the simple-evals template and reshuffled answer order; Google's is a single published number on the model card.
• Humanity's Last Exam, no tools — DaisyRose 27B 34.3 against Gemma 4 26B A4B's 8.7. This is the widest gap in either table and the one to distrust most, because the configurations differ: RadixArk ran five runs per problem with GPT-6 Luna as grader and no tools; Google's 8.7 is the no-tools figure, with 17.2 when the model is allowed a search tool.
Two rows is not a comparison, and neither vendor has run the other's harness. What the HLE gap does suggest — and it is a suggestion, not a measurement — is that a post-train aimed at long agentic sessions transfers to hard reasoning questions in a way that a sparse architecture pretrained for broad capability does not. What it does not show is that the effect is 25 points, because the harnesses are different and one of the two figures comes from a model that was given tools in its better configuration.

The rows only one side has
Both cards publish benchmarks the other does not, and the asymmetry is itself informative about what each vendor thought worth proving.
Gemma 4 26B A4B publishes a full general-capability table: MMLU Pro 82.6, AIME 2026 without tools 88.3, LiveCodeBench v6 77.1, Codeforces ELO 1718, Tau2 averaged over three 68.2, BigBench Extra Hard 64.8, MMMLU 86.3, MMMU Pro 73.8, MATH-Vision 82.4, and MRCR v2 at eight needles and 128K context 44.1. There is no Terminal Bench, no SWE-bench and no IFBench row on the card.
DaisyRose 27B publishes the agentic and instruction-following set instead: Terminal Bench 4.1 12.0 against a base of 5.6, Terminal Bench 2.1 75.5, DeepSWE v1.1 46.2, SWE Bench Pro v2 94.7 and IFBench 82.4. There is no MMLU Pro, no AIME and no Codeforces row on the card.
Read together, the two tables describe different products. Google is publishing a general model and proving it across the standard academic suite. RadixArk is publishing an agent model and proving it across the standard agentic suite. The overlap is two rows out of roughly twenty, and a buyer choosing between them is mostly choosing which suite they care about — a much simpler decision than the score columns make it look.
Where each one runs
This is the section that decides the comparison for most readers, because the deployment stories do not intersect at all.
• Gemma 4 26B A4B — hosted today at $0.06 per million input and $0.33 per million output on OrcaRouter, alongside an Obsidian-hosted build of the same architecture. Open weights are Apache 2.0, in bfloat16 (about 51.6 GB across two shards), a quantisation-aware-trained Q4_0 GGUF that puts it on a single consumer card, a w4a16 compressed-tensors build for vLLM, and a multi-token-prediction drafter for speculative decoding. Google's family post describes the two larger sizes as fitting "efficiently on a single 80GB NVIDIA H100 GPU."
• DaisyRose 27B — BF16 at about 52 GB, FP8 at about 29 GB, NVFP4 at about 23 GB and Blackwell-only, plus a DSpark speculative-decoding drafter for SGLang. RadixArk lists single-GPU deployment on H200, B200, GB300, RTX PRO 6000, RTX 5090 and DGX Spark. No hosted API price exists; third-party listings quote $0.15 per million input tokens and $1.88 per million output, which is one host's opening position rather than a vendor rate.
Those numbers put the hosted-cost gap at roughly two-and-a-half times on input and nearly six times on output, in the third party's favour. That is the honest comparison at list price. It is also the comparison that changes as soon as the workload is long: a sparse model keeps paying 3.8B of compute per token while a dense one pays 27.8B, so the gap widens in latency and in hardware-seconds long before it widens in dollars per million tokens.

What the agentic training is worth per dollar
The interesting version of this comparison is not "which scores higher" but "what does the agentic lift cost per task." RadixArk's model exists because its base model, Qwen3.8-27B, scores 5.6 on Terminal Bench 4.1 and 12.0 after post-training. If you are running a coding agent that executes commands, that is the difference between a model that mostly cannot and one that sometimes can, and a doubling of a small number is worth real money to the right buyer.
Gemma 4 26B A4B has no Terminal Bench figure at all, which makes the value claim unmeasurable from public data. What it does have is a price that makes the experiment cheap: at $0.06 per million input tokens you can run a few thousand agent trajectories for the cost of an hour of an H200, and find out whether the model fails in a way that matters to your workload. The sparse architecture makes that cheap twice over — the token price is low and the tokens come back fast.
That is the shape of the real decision. DaisyRose 27B is a specialist post-train with a specific, vendor-reported agentic claim and no hosted price. Gemma 4 26B A4B is a general model with a broad published capability table, a fast inference profile and a list price measured in cents per million tokens. If your task is long-horizon tool use and you can host, the specialist is worth benchmarking. If your task is anything else, or you cannot host, the general model at $0.06 wins on the only axis that matters.
Running both from one key
Gemma 4 26B A4B is on OrcaRouter today under google/gemma-4-26b-a4b-it, served at $0.06 per million input tokens and $0.33 per million output, with a 262,144-token context and text, image and video input. DaisyRose 27B is not on the catalogue at the time of writing and this page will not suggest otherwise — it is a self-hosted release through SGLang, vLLM and Docker.
Its base is, though. Qwen3.8-27B is served as qwen/qwen3.8-27b at $0.33 per million input tokens and $2.40 per million output, with the same 262,144-token context and the same vision tower the post-train keeps. That gives a reader two useful things: a same-family baseline to run on their own prompts before deciding whether the RL post-train is worth deploying, and a fallback route. If a self-hosted DaisyRose instance hits a sandbox timeout, a routing rule can hand the call to the base model rather than dropping it, and because we pass provider list prices through at 0% markup, a vendor-side price change on either end reaches the same key the day it lands.
The cheapest next step is not a benchmark: it is running Gemma 4 26B A4B and Qwen3.8-27B on your own agent traces, in that order — the cheap model first, to see where your tasks actually break, then the more expensive baseline to see whether the break is the base model's fault. Only after that does an eighteen-day RL post-train have anything to prove against.
Compared in this article1
Detected from this article · Benchmarks: Artificial Analysis · updated daily
