Bonsai vs Ternary Bonsai 2 27B: Two Low-Bit Bets, Two Different Rulers

Put the 1-bit Bonsai 2B vision-language model next to Ternary Bonsai 2 27B and the obvious comparison — 2 billion parameters against 27 billion, 0.43 GB of language weights against 5.93 GB — is the least interesting thing about the pair. The interesting thing is that the two models, from the same vendor and the same compression programme, do not report their quality the same way. Ternary Bonsai 2 27B reports retention: it keeps over 98% of the aggregate benchmark performance of its full-precision counterpart, measured against itself. The 1-bit Bonsai 2B reports a comparison: it achieved "comparative benchmark results" against a Qwen 3 1.7B model at 4-bit quantisation, measured against somebody else's model at a different precision. One is a statement about how much was lost. The other is a statement about how it stacks up. Neither is a statement about how good either model is in absolute terms, and the two cannot be read on the same scale.

That distinction matters more than the parameter count, because it decides what you can actually conclude from the vendor's numbers — and on the evidence published so far, the honest answer is less than the headline figures suggest.

Two quality claims, two different rulers

Ternary Bonsai 2 27B, released on 17 September 2026, is a ternary {−1, 0, +1} rebuild of Qwen3.8-27B at 1.76 effective bits per weight, and the figure PrismML leads with is that it retains more than 98% of the full-precision model's aggregate benchmark performance. That is a retention claim, and retention claims are self-referential by construction: they tell you how much of the original survived the compression, not where the original stood. A model that retains 98% of a mediocre baseline is still a mediocre model, and Qwen3.8-27B is not mediocre — but the number alone does not say so, and it cannot be compared across base models.

The 1-bit Bonsai 2B vision-language model, announced on 23 September 2026 for the Snapdragon AR1 Gen 1 glasses platform, is a 1.7B 1-bit language model plus a 0.3B 4-bit vision encoder. Its published quality statement is different in kind: PrismML evaluated it against a Qwen 3 1.7B model at 4-bit quantisation across BFCL v3, HumanEval+, MMLU Redux, IFEval, IFBench, MuSR, GSM8K and GPQA Diamond, and reported that the two configurations achieved comparative results. That is a parity claim against an outside baseline. It says the compression did not obviously cost you against the 4-bit path you would otherwise have used — which is exactly the right question for a device-class release, and a much weaker claim than "this model is good".

So there is no reading on which 98.2% beats "comparative" or the reverse. They are answers to two different questions, asked against two different references, on two different suites, by the same vendor. Anyone ranking these two models on the strength of those two sentences is ranking the sentences.

The base models come from different places

There is a second structural difference, and it is the one that will matter over the next year. Ternary Bonsai 2 27B is a re-compression of an outside model — Alibaba's Qwen3.8-27B. Its quality ceiling is set by someone else's release schedule, and its generational cadence tracks Qwen's rather than PrismML's. When Qwen ships a new 27B, the Bonsai 27B line gets a new input, and the retention figure gets recomputed against a new baseline.

The 1-bit Bonsai 2B is built on PrismML's own Bonsai 1.7B. Its ceiling is set in-house, its update cadence is PrismML's to choose, and its improvements come from the compression recipe and the kernel path rather than from an upstream model swap. That is a different kind of asset: slower to improve on raw capability, fully under the vendor's control, and — as the glasses release shows — able to be tuned for a specific accelerator in a way a general re-compression cannot.

Both are legitimate strategies. They are not interchangeable, and the one you want depends on whether your roadmap is anchored to Qwen's or to the device.

The specification contrast, one line at a time

• Parameters — a 1.7B 1-bit LLM plus a 0.3B 4-bit vision encoder in the glasses model, against a 27B-class model derived from Qwen3.8-27B in Ternary Bonsai 2 27B.

• Representation — binary {−1, +1} with FP16 group-wise scaling in the glasses model, against ternary {−1, 0, +1} with the same scaling at 1.76 effective bits per weight in the 27B.

• Language-model weights — 0.43 GB for the 1-bit 1.7B, against 5.93 GB for the PTQ1_0 packing of Ternary Bonsai 2 27B, with a 7.25 GB PQ2_0 packing also available.

• Context — 1,024 tokens on the glasses model, against a full 262,000-token context on Ternary Bonsai 2 27B.

• Accelerator — the Qualcomm Hexagon NPU through a QNN SDK with 1-bit kernel support, against Apple silicon via MLX and NVIDIA via CUDA.

• Quality claim — "comparative benchmark results" against a 4-bit Qwen 3 1.7B, against "over 98%" retention of the full-precision Qwen3.8-27B baseline. Both vendor-reported, neither independently reproduced, measured on different suites.

• Runtime — a Qualcomm NPU toolchain for the glasses model, against PrismML's own llama.cpp fork and an MLX container for the 27B, neither of which stock llama.cpp supports.

What the low-bit bet buys in each case

The compression philosophy is the same in both models and it is worth naming, because it is the family's actual thesis: the low-bit representation runs end to end — embeddings, attention, MLPs and the LM head — with FP16 group-wise scaling and no higher-precision escape hatches. Neither model keeps a float safety net for the parts that are hard to quantise. That is a more aggressive position than most quantisation work takes, and it is what makes 1.76 bits per weight and 0.43 GB weights possible at all.

Where the bet pays off differs. In Ternary Bonsai 2 27B the payoff is capability per byte on hardware you already own: a 27B-class model in 5.93 GB, running on a laptop or a phone, at 98% of its baseline. In the 1-bit Bonsai 2B the payoff is existence on a device class that had no option: a multimodal model in 0.43 GB of language weights, on an NPU, at 15.36 tokens per second. The first is a better version of something that already worked. The second is something that did not work before.

Where the tier boundary sits

These two are not alternatives, and treating them as a head-to-head misses the deployment shape both releases are pointing at. A glasses or wearable product runs the 2B locally because nothing else fits. A laptop or phone product runs the 27B because it can. The moment a product spans both — a phone app paired to glasses, a laptop app with an on-device assistant — the question stops being which model and becomes which requests each tier is allowed to answer.

That boundary is worth putting in configuration rather than in application code, because both tiers will move. The 27B line moves when Qwen ships, the 2B line moves when PrismML changes the recipe, and a hard-coded rule about context length or capability breaks on both events. A routing policy that escalates requests beyond the local tier's budget to a hosted model survives both changes; the local tier stays one tier among several rather than an assumption threaded through the client. OrcaRouter is the layer that does that escalation — one endpoint across 200-plus hosted models, failover included, with the policy expressed as configuration. Neither Bonsai is routed here; both are local downloads. What sits here is the tier above them, and with a 1,024-token local model that tier is doing most of the work.

What would settle it

Three measurements would turn this pair from two marketing claims into a comparison. A per-benchmark breakdown behind the "comparative results" line, so the trade against 4-bit is visible instead of summarised. A retention figure for the 1-bit Bonsai 2B against its own full-precision baseline — the measurement PrismML uses for the 27B line and did not publish here. And an independent evaluation of either model, since every figure in both releases currently comes from the vendor.

Until one of those exists, the defensible position is the one the numbers actually support: Ternary Bonsai 2 27B is the better model by a wide margin and the 1-bit Bonsai 2B is the only one of the two that runs on the device it was built for. Those two statements are not in tension, and the fact that the same vendor measured them with different rulers is the thing worth remembering the next time one of these figures is quoted without its baseline.