q-bonsai-27b — Bonsai-27B as a κ-object (the full 27B mind, streamed like a film)

Created using Bonsai by Prism ML.

This is prism-ml/Bonsai-27B — a 27-billion-parameter hybrid-attention model (Qwen3.5 family: 16 full-attention + 48 gated-delta linear layers) moved end-to-end into binary {−1,+1} weights (1.125 true bits/weight, embeddings and LM head included) — re-laid as a content-addressed κ-object for the Hologram serverless substrate. No re-quantization anywhere: the trained sign bits pass through byte-exact.

What's here

File	What it is
q-bonsai-27b.v1.holo	ONE Range-streamable file: boot-ordered BLAKE3-verified blocks + embedded tokenizer, lm_head packed early so a progressive reader serves the decode-critical path first.
manifest.json + b/*.gz	The same 803 blocks as loose content-addressed parts (sha256 transport axis) — the parallel-fetch fast path on CDN origins.
manifest.blake3.json + sha256-to-blake3.map.json	The canonical BLAKE3 (κ) axis.
tokenizer.gguf	The source GGUF header (tokenizer + arch), for serverless load.

Every block is verified against its content address before it is decompressed or touches the GPU. Any static file host serving these bytes is an equal origin — the bytes, not the host, carry identity.

Pins

• manifest κ: did:holo:sha256:00beea03509a79f6567879f019c04bb15dda8721f6320f8295d9ea1dd9ba54af
• root κ (tensor blocks): sha256:a5f7b8ed89546306702547e1e47541572ef58… (full value in manifest)

Architecture (from the GGUF header)

64 blocks, d=5120, ff=17408, vocab 248320; full_attention_interval=4 → 16 full-attention layers (24 heads × 256, 4 KV heads, per-head q/k RMSNorm, gated output) + 48 linear layers (gated-delta SSM: conv kernel 4, state 128, 16 groups, dt-rank 48, inner 6144); MRoPE sections [11,11,10,0], rope base 1e7; 262K trained context.

Provenance

Weights: Prism ML's Bonsai-27B (Apache-2.0) — see LICENSE and NOTICE.txt. Conversion: compile2bit.mjs q1 pass-through + holo-kappa-pack.mjs (Hologram Q substrate). Format q1: blob = [signs N·K/8 B][f32 scales N·K/128·4 B] per tensor; the GEMV kernel dequantizes inside the matmul. Browser engine support for the hybrid architecture is in active development on the Hologram substrate; the 8B sibling (HOLOGRAMTECH/q-bonsai-8b) runs live today.