Ternary Bonsai 2 27B — MLX 2-bit
Prism ML's Ternary Bonsai 2 27B checkpoint, packaged for signed-Hadamard-aware MLX inference. The original 2-bit packed weights, vision tower, tokenizer and chat template are unchanged.
| Property | Value |
|---|---|
| Architecture | Dense Qwen3.8-27B backbone, prism_hadamard_qwen35 / qwen3_5 |
| Inputs / output | Text and images / text |
| Native context | 262144 tokens; usable context depends on hardware and serving policy |
| Language weights | Affine 2-bit, group size 128, signed block-Hadamard transforms |
| Vision tower | Retained, unquantized FP16 |
| Transform arithmetic | FP32 intermediates, restored FP16 activation dtype |
| Weight payload | 8,595,477,990 bytes |
| MTP | Not included; this checkpoint has no MTP/draft heads |
Runtime support
These packed weights require their signed transforms. They must not be loaded as an ordinary affine-2-bit Qwen checkpoint or silently converted to BF16. DarkBloom's native support is under review in MLX Swift #27, MLX Swift LM #154, and d-inference #1124. Reasoning and tool calling use the native model template and serving parser. Runtime qualification, hardware-tier coverage and catalog availability are separate from artifact integrity. No production throughput guarantee is made here.
Source and license
Source: prism-ml/Ternary-Bonsai-2-27B-mlx-2bit,
revision 3f926b415992eaa2ae9dd7b573706494d6bbf787.
Weight SHA-256: 130de5925082c168b7866b2e91b52e44abbafc99017e3ca352b77b5b55a269ed.
See LICENSE, NOTICE.txt and UPSTREAM-README.md for the original attribution
and model documentation. Upstream benchmark claims describe their own runtime
and methodology, not DarkBloom measurements. artifact-provenance.json records
the unchanged payload identity. The older bundled text-only runtime notes do
not describe the current schema-2 vision-capable pack; use a compatible loader.