PrismML Bonsai

PrismML Bonsai is a family of open-weight 1-bit and ternary language models, released under Apache 2.0, that shrink 27B-class AI to a few gigabytes so it runs locally on laptops, workstations and phones. It is for enterprises and developers who want private, low-cost, offline inference without sending data to a cloud API.

Bonsai is a family of extremely compressed open-weight language models from PrismML, a Pasadena startup built on Caltech research and led by CEO Babak Hassibi, a Caltech professor. Instead of the 16-bit weights most models ship with, Bonsai stores weights as 1-bit {-1, +1} or ternary {-1, 0, +1} values with group-wise scaling, applied across embeddings, attention, MLP layers and the LM head. PrismML emerged from stealth on March 31, 2026 with 1-bit Bonsai 8B, 4B and 1.7B (the 8B fits in about 1.15 GB and averaged 70.5 across its benchmark suite), then added Ternary Bonsai at 1.58 bits, Bonsai Image 4B for image generation, and in July a 1-bit Bonsai 27B small enough to run on a phone. On September 17, 2026 it released Ternary Bonsai 2 27B, a rebuild of Qwen3.8 27B at 1.76 effective bits per weight that occupies 5.9 GB, more than 9x smaller than full precision, and retains a claimed 98.2% of the base model's aggregate benchmark score (83.9 versus 85.4). It accepts text and image input, has a 262K-token context window, reaches up to 143 tokens per second on an RTX 5090 and 46.8 on an Apple M5 Max, and is distributed on Hugging Face as GGUF and MLX builds under Apache 2.0. Running Bonsai 2 currently requires PrismML's forks of llama.cpp and MLX because the necessary kernels are not upstream yet. The company has raised $22.25 million in seed funding from Khosla Ventures, Cerberus and Caltech, and offers enterprises custom domain-specific post-training and hardware inference optimization. TechCrunch names Multiverse Computing as its closest competitor; the practical alternative for most teams is conventional 4-bit quantization of the same open models.

Category

AI Models & APIs

Pricing

Free, Contact for pricing

Target Market

CTOs, AI/ML Engineers, Enterprise Developers, Device and OEM Product Teams

Deployment

Open-source, Self-hosted, Edge-first

Headquarters

Pasadena, California, USA

• ✓
  End-to-end 1-bit and ternary weights

  Every layer, including embeddings and the LM head, uses 1-bit or ternary weights, cutting memory roughly 9-14x versus 16-bit models.

• ✓
  Bonsai 2 27B near-lossless compression

  Ternary rebuild of Qwen3.8 27B at 1.76 bits per weight in 5.9 GB, retaining a claimed 98.2% of aggregate benchmark performance.

• ✓
  Multimodal, long-context input

  Bonsai 2 27B accepts text and images with a 262K-token context window, enabling document and screenshot analysis on local hardware.

• ✓
  Broad hardware support

  Runs on CUDA, Metal, Vulkan, ROCm and CPU backends, and the 1-bit 8B runs natively on iPhone via MLX Swift at about 44 tokens per second.

• ✓
  Apache 2.0 open weights

  Models are free to download, modify and deploy commercially from Hugging Face in GGUF and MLX formats, with no per-token fees.

• ✓
  OpenAI-compatible local serving

  Documentation covers llama.cpp, MLX, an OpenAI-compatible server and Open WebUI, so existing client code can point at a local endpoint.

• ✓
  Enterprise post-training and optimization

  PrismML offers custom domain-specific post-training and hardware inference optimization for companies building Bonsai into their own products.

• •
  Private on-device assistants

  Enterprises run a 27B-class assistant entirely on employee laptops, keeping sensitive documents off third-party cloud APIs and eliminating per-token costs.

• •
  Edge and offline deployments

  Field, factory or secure-facility teams deploy capable language models on disconnected hardware where cloud inference is unavailable or prohibited.

• •
  Lower-cost self-hosted inference

  Platform teams fit a 27B-class model on a single GPU with room left for long context, one tester reporting about 90K tokens of context on 20 GB of VRAM.

• •
  On-device AI in mobile apps

  Mobile developers ship on-device text and image understanding in iOS apps using MLX builds, avoiding server round-trips, latency and per-call fees.

• •
  Zero-install browser inference

  Teams run Ternary Bonsai 2 27B in the browser via WebGPU for internal tools and demos, with no server or local installation required.

All Bonsai models are free to download and use commercially under Apache 2.0, and there is no hosted API or per-token pricing. The real cost is your own hardware and engineering time, including running PrismML's llama.cpp and MLX forks. Custom domain-specific post-training and hardware inference optimization are sold to enterprises on unpublished terms.

This page was written from 11 sources, 7 on domains other than prismml.com.

1. 1.prismml.com — prismml.comvendor
2. 2.prismml.com — bonsai 2 27bvendor
3. 3.prismml.com — bonsai 8bvendor
4. 4.prismml.com — prismml launches worlds first 1 bit ai modelvendor
5. 5.docs.prismml.com — introduction
6. 6.techcrunch.com — prismml hopes its tiny llm could change how we all use ai
7. 7.siliconangle.com — prismml launches bonsai 2 27b a high intelligence ai model s
8. 8.hn.algolia.com — 49746618
9. 9.hn.algolia.com — search
10. 10.github.com — Bonsai demo
11. 11.huggingface.co — bonsai 2