PrismML releases three Bonsai 2 ternary-weight models; 9B model compressed to 2GB
BIBIBI
AT A GLANCE
PrismML released three Bonsai 2 models, saying the 9B version was compressed to 2GB, and can generate up to 34 tokens per second locally on an iPhone.Token。
Article
PrismML compresses 9000000000 the weights of the parameter model to 2GB: generates locally on iPhone 34 tokens Token Beating AI Flash: U.S.-based AI company PrismML has added to the Bonsai 2 family 2B、4B、9B three new models, all based on Alibaba's Qwen3.5. The team constrained the language model's main weights to -1、0、+1 three possible values, reducing the model files to 0.43GB、0.95GB and 2GB, approximately one-ninth the size of the original 16-bit weights. The new models still support inference, image understanding, and tool calling, and are primarily intended for local Agents on devices such as phones and personal computers.
PrismML said that on the iPhone 18 Pro Max, the 9B model generates up to 34 Token，2B reaches 120 tokens per second.
In official testing, the 9B model's coding and instruction-following capabilities nearly matched those of the original model of the same size,Qwen3.5 but the gap was clear on complex tasks. In the multi-turn dialogue and tool-use Tau2Bench benchmark, Bonsai 2 9B scored 73.02, while the original reached 85.25; scores for complex document parsing were 73.8, while the original reached 88.73.
The three models have released their weights under the Apache 2.0 protocol.
Original link https://m.theblockbeats.info/flash/372238
Key points
01
PrismML has added to the Bonsai 2 family 2B、4B、9B three new models, which the article says are all based on Alibaba's Qwen3.5。
02
The team constrained the language model's main weights to -1、0、+1 three possible values, reducing the three model files to 0.43GB、0.95GB and 2GB。
03
The article says the model size is approximately one-ninth that of the original 16-bit weights.
04
The new models support inference, image understanding, and tool calling, and are primarily intended for local Agents on devices such as phones and personal computers.
05
PrismML said that on the iPhone 18 Pro Max, the 9B model generates up to 34 tokens per second, while the Token，2B model reaches 120 tokens.Token。
06
In official testing, Bonsai 2 9B scored Tau2Bench 73.02, while the original scored 85.25; the complex document parsing score was 73.8, while the original scored 88.73。
07
The three models have released their weights under the Apache 2.0 protocol.
AI-assisted interpretation
The following is analysis, separate from reported facts. Verify important claims independently.
This technology constrains the model's main weights to three possible values to reduce the model file size, making it easier to run the model directly on a phone or personal computer. Smaller files and faster generation benefit local Agents, but official testing also showed that the 9B model still lagged behind the original in multi-turn dialogue, tool use, and complex document parsing.
Why it matters to readers
If the file size, on-device speed, and functional performance are independently verified, this compression method could lower the storage and compute barriers to running models on personal devices.
Beginners can understand it as “compressing a larger AI model and putting it on a phone to run.” The key is not just how many Token are generated per second, but also the loss of capability in handling complex tasks after compression.
Risks and unknowns
The on-device generation speed is claimed by PrismML; the original text provides no independent verification results.
The capability scores come from official testing; the original text does not specify the complete test conditions or reproducibility details.
The specific model version referred to by the original text as “the original Qwen3.5” is not clear.
The source text does not provide actual memory usage, power consumption, or performance on different devices.
The availability of the weights as open is stated only in the source text; no repository information is provided in the evidence for verification.
Related Developments
Loading event timeline…
Related concepts
Model weights
This term is not in the glossary yet. Browse related concepts in the glossary.
