What happened
Prism ML’s Hugging Face repository lists Ternary-Bonsai-2-9B-gguf as an image-text-to-text model. Its page describes it with the terms “GGUF,” “llama.cpp,” “ternary,” and “on-device.” The repository includes multiple GGUF files, among them Q2_0, PTQ1_0, PQ2_0, and Q8_0 variants, as well as a DFlash2 Q8_0 file and two multimodal projector files, one in BF16 and one in Q8_0. The supplied release description reports a 9B model and an approximately 2.42 GB ternary-weight artifact.
The model page provides instructions for use through llama.cpp. It shows an installation command for macOS and Linux using the llama.app installer, followed by commands to start a local OpenAI-compatible server with a web UI or run inference in a terminal. The example selects the Q8_0 variant. For Windows, the page gives a WinGet command to install llama.cpp and likewise shows a local server command. These are the repository’s documented routes; the supplied material does not say that all listed variants work identically across environments.
The release is not presented as a hosted service with a stated usage price. The supplied page excerpt shows a model repository, downloadable files, and local-app instructions; it does not provide a price, an inference-provider listing, or a claim that the model is available through a particular commercial API. The repository metadata identifies the license as Apache-2.0. The excerpt lists a total model-file size of about 1.57 GB for one file set, while the release description gives approximately 2.42 GB for the ternary-weight model. These refer to different displayed metadata and should not be treated as interchangeable measures of every configuration’s disk or runtime footprint.
The model card’s excerpt does not provide benchmark results, supported image resolutions, latency figures, memory requirements, or a comparison against the earlier Bonsai versions. It also does not specify which hardware the ternary kernels require. The page’s tags describe the project as multimodal and local-oriented, but tags and sample commands are not evidence of performance on a particular device.
What the archive adds
This is a new release in a story the archive has covered before. PrismML — Introducing Ternary Bonsai 2 2B, 4B, and 9B: Bringing Local Agents to Everyday Devices covered the Bonsai family and the trade-off between compact local deployment and capability retention. The new repository adds a specific GGUF packaging of the 9B vision-language model, with multiple quantized file options and local llama.cpp instructions; it does not, in the supplied evidence, establish how its capabilities compare with the earlier models.
Trend signal
This release supports agent economics moves to completed work only weakly: compact weights may change the practical cost and hardware choices for local use, but the source gives no cost-per-task or verified-completion data. More directly, the archive’s prior Bonsai coverage treats local deployment as a memory-and-power allocation decision, while noting that capability retention varies by task. This release adds another packaging option, not evidence that the underlying trade-off has been resolved.
Why it matters
For teams evaluating local multimodal systems, a quantized GGUF artifact with llama.cpp instructions lowers the first-step friction: builders can inspect the available variants and try the documented local server or CLI route without treating this as an API-only model. The approximately 2.42 GB figure is a useful signal for comparing artifact sizes, but it is not a complete hardware budget. Runtime memory also depends on the selected files, context, image handling, and implementation; the supplied source does not quantify those costs.
The important engineering question is therefore not simply whether a 9B model can be stored locally. It is whether a specific variant can reliably handle a team’s image-and-text tasks at acceptable latency on its target hardware. Builders should test representative inputs, record the exact file variant and runtime, and validate outputs before assigning the model a role in a workflow. The repository establishes a downloadable option and a launch path, not production suitability.
Limits and uncertainty
The supplied model-page material does not establish accuracy, benchmark performance, supported image sizes, inference speed, or minimum system memory. It does not explain the quality differences among the listed GGUF variants or say whether the reported approximate size applies to a complete deployment including the multimodal projector. The page’s commands demonstrate an intended local path, but do not prove performance on every platform or that upstream llama.cpp supports all ternary-weight kernels. The separate discovery note says ternary kernels need PrismML’s llama.cpp fork while Q2_0 runs upstream; that technical detail is not supported by the primary source excerpt provided here, so this note does not treat it as a verified claim. Teams should confirm compatibility in the repository and test the precise configuration they intend to deploy.
Reader takeaway
Prism ML has made a 9B vision-language model available as GGUF files, including a reported roughly 2.42 GB ternary-weight option, with local llama.cpp usage instructions. That makes it worth evaluating for local multimodal experiments, but the release page alone does not show how well it performs or what hardware it needs. Treat the artifact size as a starting point for testing, not as proof of an inexpensive or production-ready deployment.