What happened
Prism ML’s Hugging Face model card describes Ternary-Bonsai-2-2B as a vision-language model derived from Qwen3.5-2B, with the base architecture unchanged. The language model has approximately 1.88 billion parameters across 24 blocks, with a hidden size of 2,048 and tied embeddings and language-model head. The separate vision tower has approximately 0.33 billion parameters and 24 blocks. The listed context length is 262,144 tokens, inherited from the base model’s hybrid-attention architecture, which combines 18 linear-attention and six full-attention layers.
The release’s defining choice is ternary language weights: values of −1, 0, or +1, with group-wise FP16 scales. The model card lists three packings of the same weights. PTQ1_0 uses dense trits and is 0.43 GB; PQ2_0 uses 2-bit slots and is 0.52 GB; Q2_0 is 0.55 GB. The card compares these with about 3.8 GB at FP16. These figures describe the language-model files; image input uses a separate vision-projector file, available as Q8_0 at 0.36 GB or BF16 at 0.67 GB. Prism ML recommends Q8_0 for the projector.
Runtime compatibility depends on the packing. The PrismML llama.cpp fork runs all listed files. The model card says upstream llama.cpp can run Q2_0 from release b11556 onward, which added the required Hadamard activation runtime; older builds may produce incorrect output without warning. Upstream llama.cpp does not support PTQ1_0 or PQ2_0 in the described release path. The card also gives local commands for text and image input and describes an OpenAI-compatible local server. It lists integrations and instructions for tools including Ollama, LM Studio, vLLM, Docker Model Runner and MLX; the MLX variant is identified separately as Ternary-Bonsai-2-2B-mlx-2bit and requires mlx-vlm 0.7.8 or newer.
The model is released under Apache 2.0. The Hugging Face page lists 873,081 downloads in the last month. Prism ML’s card recommends different sampling settings for thinking and non-thinking modes, warns against greedy decoding with thinking enabled, and says to allow at least 8,192 output tokens when thinking is on. It also reports that speculative decoding with a draft head was slower than plain decoding on the team’s everyday prompts, so the release does not include a draft model.
What the archive adds
New Runtime’s earlier coverage described Prism ML’s 9B Ternary-Bonsai GGUF release as a 9B VLM available at about 2.42 GB (earlier coverage, November 5, 2026). The 2B release is a new event in the same model family, not a restatement: it adds a smaller model size and documents three language-weight packings, a separate vision projector, and a compatibility path for upstream llama.cpp.
Trend signal
This does not yet fit a pattern New Runtime tracks. The release shows compact local model distribution and runtime-specific packaging, but a single model release does not establish a broader shift in deployment economics, agent interfaces, or interoperability.
Why now
Prism ML’s model card presents the release as a runnable local option and distinguishes the packing that works with upstream llama.cpp from those requiring its fork. That makes runtime compatibility central to using the files today: builders can choose between smaller packings with a fork and the Q2_0 option supported by a sufficiently recent upstream build.
Why it matters
For teams evaluating local image-and-text processing, the release makes the model files and deployment path concrete enough to try without first designing a hosted API integration. The language-weight files are substantially smaller than the card’s FP16 comparison, but the vision tower is a separate download and adds its own memory and runtime requirements. A small checkpoint is not by itself evidence that a model will meet latency, quality, or hardware constraints in production.
The distinction between packings also matters operationally. Choosing the smallest artifact may mean adopting Prism ML’s llama.cpp fork, while choosing Q2_0 allows use of upstream llama.cpp at the stated minimum version. That creates a useful evaluation question for builders: whether file-size savings justify carrying a specialized runtime, or whether a more widely supported packing is preferable. Teams should test representative images, prompts, concurrency, and output quality on their target devices before treating the release as a production substitute.
Limits and uncertainty
The model card’s parameter counts, file sizes, recommendations, and performance observations are publisher claims, not an independent evaluation. It does not establish comparative image-understanding quality, latency, memory use under a team’s workload, or hardware coverage beyond the listed backends. The reported speculative-decoding result is based on Prism ML’s own tests, with no benchmark protocol or broader workload results in the excerpted material. Download counts also do not demonstrate active use or production adoption.
Skill to build now
Build a small local VLM evaluation harness around the exact runtime you expect to deploy. Compare Q2_0 on upstream llama.cpp b11556 or newer against PQ2_0 on the PrismML fork, using the same set of representative images and prompts. Record startup time, peak memory, response latency, and whether outputs meet a human-defined quality bar; include the separate Q8_0 vision projector in your measurements.
Reader takeaway
Ternary-Bonsai-2-2B gives builders a compact local VLM to evaluate, with explicit choices between three language-weight packings and a separate vision file. The important practical detail is that the smallest files require Prism ML’s llama.cpp fork, while upstream compatibility is documented for Q2_0 on a recent build. Treat it as an evaluation candidate, not a proven production model.