PrismML has introduced Ternary Bonsai 2, a family of 2B, 4B and 9B vision-language models built on Qwen3.5 where every weight is stored as one of just three values. The files are tiny, the claims are bold, and the numbers deserve a careful read. Here is what the company says, and what I would verify before putting one on a phone.
Ternary weights in plain terms
Normally a model weight is a decimal number, stored with 16 bits or, after quantization, 8 or 4 bits. A ternary weight can only be -1, 0 or +1. That is the whole vocabulary. PrismML’s announcement says the new models use exactly this representation.
Picture a huge grid of dials. Instead of letting each dial sit anywhere along a smooth scale, you only allow it three positions: reverse, off, or forward. Multiplying by -1, 0 or +1 is far cheaper than multiplying by arbitrary decimals, and storing three choices takes a lot less room than storing a precise number. That is where the small files come from.
The trick is keeping the model smart after that much compression, and that is the part PrismML says it has worked on.
What PrismML reports
According to PrismML, the family includes three models based on Qwen3.5 at 2B, 4B and 9B, with footprints of 0.43 GB, 0.95 GB and 2 GB. The company says that is more than 9 times smaller than the full-precision counterparts. The weights are available under the Apache 2.0 License.
On quality, PrismML evaluated across 19 benchmarks in six skill areas and publishes a table comparing each Bonsai model against its Qwen3.5 baseline. For the 9B, its per-area retention figures range from about 92% to 100%, covering agentic tool calling, math, instruction following, coding, knowledge and reasoning, and vision. That lines up with the claim that the 9B keeps more than 90% of Qwen3.5-9B’s scores. The table has weaker spots too. For example, PrismML lists the 9B at 84.7% retention on GPQA Diamond, and the smaller models drop further on some knowledge and vision tests. I would read the table row by row for the skills your app actually needs.
On speed, PrismML says Bonsai 2 models deliver up to 5x the throughput and 3x the energy efficiency of full-precision counterparts on hardware such as an RTX 4090 and an M5 Pro. On an iPhone 18 Pro Max, the company quotes up to 34 tokens per second for the 9B and up to 120 for the 2B. The company also says the 9B uses roughly 40% fewer tokens on standard instruction-following and everyday math tasks, which would help battery life on its own.
For platforms, PrismML says the models run on NVIDIA GPUs via llama.cpp and on Apple devices through MLX.
A runtime note for Ollama users
If you live in Ollama, the v0.40.6 release includes a change titled “gguf: add the q2_0 tensor and file types.” That is groundwork for GGUF files that use the q2_0 format. The release page does not mention Bonsai, and PrismML’s announcement does not name Ollama, so I cannot tell you that these models load there. Treat it as a promising sign and test it with your own pull of the files.
Caveats worth keeping in mind
- Every figure above is PrismML’s. I have not seen independent phone measurements, and “up to” numbers describe a best case on one device.
- The retention percentages compare against Qwen3.5 baselines that PrismML ran itself.
- Runtime support decides everything with exotic formats. Check that the exact file format you download is supported by the exact llama.cpp, MLX or other build you plan to ship.
- Some rows in the table are marked N/A for the smaller models, so not every benchmark exists for every size.
Where this fits
Last week I wrote about small active models and local agents, and Bonsai 2 adds another option to that picture: take a model you already know and shrink it aggressively. A 2 GB vision-language model with tool calling that fits comfortably on a phone opens up plenty of app ideas, such as reading a document on device and handing the result to your app’s own tools without a network round trip.
If I were evaluating it, I would pick the size that matches my memory budget, run my real prompts and a few image inputs against both the Bonsai model and its Qwen3.5 baseline, and compare quality, speed and battery on the actual device. Ten minutes of that tells you more than any launch table.
Small files are exciting, and the details matter. Before you build on any of this, read the model cards, the whitepaper and the runtime docs closely, so you know exactly which formats, hardware and licenses you are working with.