Advertise here
#160

DaisyRose 27B

Radixark Release: 2026-11-18 Tested on: 2026-11-22 00:26 radixark/daisyrose-27b::xhigh
27.8BDenseOpen source

Summary

DaisyRose 27B scores 8.1 on AI BENCHY and ranks #160. It has 9.7 reliability, a 68.1% pass rate, $1.559 total cost, and 111.38s average response time.

What makes DaisyRose 27B unique: It uses unusually many reasoning tokens, which can help explain its slower or more expensive runs.

Model facts

Researched on 2026-11-22

Reported
Parameters
27.8B
Architecture
Dense
Availability
Open source
License
Apache-2.0

RL-only post-train of Qwen3.8-27B (hybrid Gated DeltaNet). The card lists 27.8B total and the same activated count, so it is dense. Public Apache-2.0 weights with SGLang, vLLM and transformers instructions; the RL task set is not released. Released 2026-11-18.

Score

8.1

Consistency

8.9

Total Output Tokens

799,132

Total Input Tokens

401,324

Input Price

$0.150 / 1M

Output Price

$1.875 / 1M

Cache Read Price

$0.038 / 1M

Cache Write Price

N/A

Cache prices apply to input tokens. Reads reuse cached prompts; writes store them and can cost extra. Output tokens use the output price.

Tests Correct

Wrong Tests: 9

Attempt pass rate: 68.1%

Flaky tests

3

Flaky tests had mixed outcomes across runs (at least one pass and one fail).

Response Time (avg)

111.38s

Response Time (max): 960.01s

Response Time (total): 2561.63s

Hamster playing table tennis

Prompt: Create a detailed SVG illustration of a hamster playing table tennis.

#160 DaisyRose 27B

xhigh
Cost
$0.063
Time
189.8s
Tokens
33,346 tok

Charts

Choose the first model, then click a second model to open a side-by-side page.

Total Output Tokens

Score vs Total Output Tokens

Quick Compare

Category Breakdown

Category Score Consistency Tests Correct
Agentic 10.0 10.0
Anti-AI Tricks 10.0 10.0
Coding 9.9 10.0
Combined 7.2 5.8
Data parsing and extraction 7.3 5.8
Domain specific 3.0 10.0
General Intelligence 5.1 10.0
Instructions following 9.8 10.0
Puzzle Solving 5.5 7.1
Tool Calling 10.0 10.0
Trivia 3.0 10.0

Compared models