
DaisyRose 27B vs Bonsai 27B: Same Size, Opposite Bets
- 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 Bonsai 27B are both twenty-seven-billion-parameter models with a 262,000-token context, an Apache 2.0 licence and a Qwen base, and they are close to opposites in every other respect. DaisyRose 27B, published by RadixArk on 18 November 2026, spends eighteen days of H200 time and about 63,000 GPU-hours on reinforcement learning to raise what a 27B can do on agentic tasks; it ships as 52 GB of BF16 weights and wants a datacentre GPU. Bonsai 27B, published by PrismML on 14 July 2026, spends its effort in the opposite direction: it takes a 27B and compresses the weights to 1.125 bits each so the whole thing fits in 3.8 GB and runs on a phone. Neither is a better model. They are two different answers to the question "what should a 27B cost," and the answer you want depends entirely on whether the constraint is quality or bytes.
One caveat before any comparison, because it decides how much of this page can be arithmetic: the two vendors do not share a single benchmark. RadixArk reports Terminal Bench, DeepSWE, SWE Bench Pro, HLE, IFBench and GPQA Diamond. PrismML reports a fifteen-benchmark suite of its own plus a derived "intelligence density" figure. There is no row on which these two models have both been measured, so nothing below subtracts one score from another.
What each one is built from
The lineage is the first real difference, and it is easy to miss because both say "Qwen 27B."
• DaisyRose 27B — a post-train of Qwen3.8-27B, architecture untouched. Same tokenizer, same chat template, same vision tower, same 262,144-token context. RadixArk's training changed the behaviour and nothing else, and it describes the result as a drop-in replacement for the base.
• Bonsai 27B — a low-bit rebuild of Qwen3.6-27B, one generation older. It is not a post-train: the language network is re-encoded end to end at 1-bit or ternary precision, embeddings, attention, MLP projections and LM head alike, with no higher-precision escape hatch anywhere in the backbone.
That generational gap is a real handicap for Bonsai and it is rarely stated plainly. Qwen3.6-27B is the model RadixArk chose to improve on; Bonsai 27B is a compressed copy of it. So a reader comparing the two is not only comparing a post-train against a compression, but a post-train of a newer base against a compression of an older one. Two changes in one step, and neither vendor's table separates them.
Two packs, two footprints
Bonsai 27B ships as two weight formats of the same model, and the choice between them is the only real decision inside that release.
• Binary pack — {−1, +1} weights with FP16 group-wise scaling, packed as GGUF Q1_0_g128 at 1.125 effective bits per weight. The shipped file is 3.803 GB; the card rounds the deployed footprint to about 3.9 GB.
• Ternary pack — {−1, 0, +1} weights at the same scaling. A ternary value carries log₂3 ≈ 1.585 bits, so the representation's floor is about 1.71 bits per weight, or 5.9 GB ideal. The shipped file is 7.165 GB, because current kernels store each ternary value in a two-bit slot — 2.125 bits per weight deployed. That gap is a kernel limitation, not a modelling choice, and PrismML publishes both numbers.
DaisyRose 27B ships in the ordinary way: BF16 at about 52 GB, FP8 at about 29 GB, NVFP4 at about 23 GB and Blackwell-only. Against Bonsai's 3.803 GB that is a factor of roughly fourteen at the binary end, and the compression is not free — PrismML's own suite puts the binary pack at 89.5% of its full-precision baseline and the ternary pack at 94.6%.
Where the compression is actually paid for
PrismML's category breakdown is the most useful thing in either release, because it shows the loss is not spread evenly. Against its own FP16 Qwen3.6-27B baseline at 85.07 on a fifteen-benchmark thinking-mode suite:
• Binary pack — 76.11 overall, 89.5% retention. Math falls from 95.33 to 91.66 and coding from 88.74 to 81.88. Instruction following falls from 78.47 to 65.74, agentic and tool-calling work from 80.00 to 66.03, and vision from 72.61 to 59.57.
• Ternary pack — 80.49 overall, 94.6% retention, recovering about half the distance to full precision for 3.3 GB more weight.
Read that list against what DaisyRose 27B was trained for and the two releases stop being comparable in any simple sense. Bonsai 27B's weakest category after compression is agentic and tool-calling at 66.03 against a baseline of 80.00 — a fifteen-point loss in exactly the area RadixArk spent eighteen days of GPU time improving. One model pays for its smallness in the currency the other model spends its compute buying. If your workload is a coding agent that calls tools in a loop, the binary pack of Bonsai 27B is the worst-fit artifact in this comparison and DaisyRose 27B is the best, and nothing else in either table changes that ordering.

What each one asks of your hardware
This is where the comparison becomes concrete rather than rhetorical.
• DaisyRose 27B — a single H200, B200 or GB300, or an RTX PRO 6000, an RTX 5090 or a DGX Spark at NVFP4. The recommended serving path is SGLang with a DSpark speculative-decoding drafter, four data-parallel replicas across a four-GPU node at tensor parallelism 1 with an FP8 KV cache. RadixArk publishes no tokens-per-second figure for any of those configurations.
• Bonsai 27B — llama.cpp on CUDA, Metal or CPU, with the 1-bit and ternary kernels living in PrismML's own fork rather than upstream, so stock llama.cpp will not load them correctly yet. The binary pack fits a 12 GB phone; the ternary pack needs a little more headroom. Add roughly 2.4 GB if you take the drafter and the vision tower as well, because the 3.9 GB headline is the language weights and not the deployment.
There is no configuration in which these two compete for the same slot. A 3.8 GB binary model on a phone and a 52 GB BF16 model on an H200 do not trade off against one another; they serve different products. The overlap is narrower and more interesting: an NVFP4 DaisyRose at 23 GB against a ternary Bonsai at 7.2 GB, both of which fit a single consumer card, is the only place where the choice is genuinely a trade rather than a category.
Evidence, and how much of it there is
Bonsai 27B has four months of downloads behind it — the ternary GGUF repository passed 590,000 in October 2026 against just over 400,000 for the binary pack, with the MLX containers for both above a million each. What that suggests is that most of this model is being run on Apple hardware rather than on servers, and that the people downloading it are buying the extra five points of retention rather than the smallest file. It is usage evidence rather than quality evidence, but it is evidence of a kind neither vendor's benchmark table provides.
DaisyRose 27B has three days and no independent run. Its seven benchmark rows are RadixArk's, measured on RadixArk's harnesses, at temperature 1.0, top_p 0.95, top_k 20 and reasoning effort xhigh, and published as the better of its BF16 and FP8 runs. Its cross-model comparison against larger systems draws on other parties' published figures rather than on side-by-side runs, and the card says so.
Neither model has been on the same harness as the other. Any page that puts a DaisyRose number next to a Bonsai number and draws a conclusion is inventing a comparison that nobody has run.

Which one to pick
The decision resolves on one question, and it is not a benchmark question: does the model have to run where there is no datacentre?
If yes, the answer is Bonsai 27B, and specifically the ternary pack unless the device is genuinely memory-starved. The binary pack's fifteen-point loss on agentic and tool-calling work is the single largest quality concession in either release, and the 3.3 GB that the ternary pack costs buys back half of the distance to full precision. On a 12 GB phone there is no alternative; on anything with 8 GB to spare for weights, the ternary pack is the better artifact.
If no — if the model runs behind an API or on a serving node — Bonsai 27B's central advantage evaporates and DaisyRose 27B's central advantage is the only one left standing. A model that fits on an H200 is not a model you need compressed, and what you would be buying with Bonsai in that slot is a 27B capped at 89.5% or 94.6% of a one-generation-older base. RadixArk's release is aimed at exactly that reader, and the claim to test is whether the Terminal Bench 4.1 jump from 5.6 to 12.0 reproduces on a public harness.
A cheap way to run the comparison yourself
Neither DaisyRose 27B nor Bonsai 27B is on OrcaRouter's catalogue at the time of writing, and this page will not pretend otherwise. Their shared ancestor is: 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 vision tower that DaisyRose 27B keeps. Running the base on your own prompts is the cheapest way to establish what the post-train is being asked to improve, and it costs a few cents rather than a rented H200.
The same endpoint is a useful safety net afterwards. If a self-hosted DaisyRose deployment hits a full context window or a broken sandbox mid-task, an automatic failover rule sends the request to the base model instead of returning an error — a degraded answer rather than a failed call. That is the practical version of the trade this whole comparison is about: the ceiling is worth paying for, but only if there is something underneath it.
What would change this page is a single shared harness. If PrismML compresses Qwen3.8-27B — the base DaisyRose improved — the two models would finally be measurable against each other, and the question "what does a 27B lose by becoming 3.8 GB" would have a number instead of a category.
