DaisyRose-27B for hipfire
Text-only RadixArk/DaisyRose-27B artifacts for hipfire, the Rust-native LLM inference engine for AMD GPUs.
DaisyRose-27B is RadixArk's RL post-train of Qwen3.8-27B for long-horizon agent
tasks (blog). It keeps the
Qwen3.8-27B architecture (qwen3_5 dense hybrid: DeltaNet + full attention,
64 layers, 1 MTP layer), so it runs on hipfire's Qwen3.5-family path unchanged.
Files
| File | What | Size | sha256 |
|---|---|---|---|
daisyrose-27b.mq4 |
MQ4V2 base tier (embeddings + lm_head at Q8, norms F16) | 15.66 GB | a380a49a6bc6b7cbe8d0aaf61147f943ecdc91c17b05af81402ff5358e14aeb3 |
daisyrose-27b.mtp |
MTP speculative-decode head (MQ4G256) | 215 MiB | cc2a02347b3319477f48ece45be9b84922a93ec886bf49be775630e00bf8adde |
Keep both files in the same directory: the daemon attaches <model>.mtp
automatically when MTP is enabled.
Not AWQ-calibrated. daisyrose-27b.mq4 is the uncalibrated MQ4V2 base tier
(hipfire-quantize --format mq4v2 --tier base). Unlike
qwen3.8-27b.mq4, it carries
no AWQ activation-scale sidecars, so expect somewhat higher KLD than that file at
the same size. A calibrated build may replace it later under the same name.
Measured (R9700, gfx1201)
Greedy, kv q8, one daemon session, prompts: factual / trick-question reasoning /
short Python function / ~120-word prose.
| Mode | Decode |
|---|---|
| AR | ~34 tok/s |
| MTP K=3 (Ο β 2.0β2.3) | ~60 tok/s |
MTP output was token-identical to AR on every prompt. Attractor checks (unique-token ratio, max-token frequency on first/last 128 tokens, 3-gram repetition) were clean on all rows. Like Qwen3.8, the model reasons in a think block first; length-constrained writing tasks can spend a large budget there counting words, so give it room.
Usage
hipfire pull daisyrose:27b
hipfire run daisyrose:27b "The capital of France is"
License and attribution
Apache-2.0, inherited from RadixArk/DaisyRose-27B
(itself a finetune of Qwen/Qwen3.8-27B). These files are a quantization of the
upstream BF16 weights; the vision tower is omitted. All credit for the model goes
to RadixArk and the Qwen team.