Prism ML

Ternary Bonsai 27B

Frontier
611.7KDownloads1.0KLikesJul 2026Veröffentlicht262K TokenKontextApache 2.0Lizenz87 StarkQualität

Ternary Bonsai 27B (27B parameters) requires approximately 9.7 GB of VRAM with Q2_0_G128 quantization. For the best balance of quality and speed, we recommend hardware with at least 12 GB of VRAM.

Loslegen

— kopieren & einfügen, um lokal auszuführen

Copy-paste commands to run Ternary Bonsai 27B on your machine.

Run

docker run --rm -it ghcr.io/ggerganov/llama.cpp:full \ --hf-repo "prism-ml/Ternary-Bonsai-27B-gguf" \ --hf-file "Ternary-Bonsai-27B-gguf-Q2_0_G128.gguf" \ -c 4096 -ngl 99

Quick specs

Parameters27B
Architecturedense
Context262K tokens
Modalitytext+vision
Min RAM7.2 GB
Rec. RAM7.2 GB (Q2_0_G128)
LicenseApache 2.0
FamilyBonsai
Code Chat Reasoning

About this model

Ternary Bonsai 27B is Prism ML's ternary-weight build of Qwen3.6-27B, running full 27B-class reasoning at a true 1.71 bits per weight. It deploys in roughly 7.2 GB instead of ~54 GB at FP16 while retaining about 95% of full-precision quality, and keeps the 262K context practical via the Qwen3.6 hybrid-attention backbone and 4-bit KV-cache quantization.

  • ~7.2 GB deployed footprint for a 27B model — about 9.4x smaller than FP16.
  • Retains ~95% of FP16 quality: 80.49 average across 15 thinking-mode benchmarks.
  • True 1.71 bits/weight end-to-end — embeddings, attention, MLP and LM head, with no high-precision escape hatches.
  • Scores above a conventional IQ2_XXS build (72.73) at under two-thirds of its footprint.
  • ~26 tok/s on an Apple M5 Pro laptop; ships a DSpark drafter for a 1.34x CUDA speedup.

Verwandte Modelle

Deine Hardware

Erkennung...

Schnellauswahl

Beste Hardware

Top-Empfehlungen für Ternary Bonsai 27B

Dieses Modell ausführen

Quantisierungsoptionen

VRAM-Schätzungen nach Quantisierungsstufe

No hardware detected — fit column shows raw VRAM estimates

QuantBitsVRAMQualityFit
Q1_0_G128
1.125
3.9 GB
Very Low
Q2_0_G128
1.71
7.2 GB
Low
Q2_K
2
10.5 GB
Low
Q3_K_S
3
13.2 GB
Low
NVFP4
4
15.1 GB
Medium
Q4_K_M
4
16.5 GB
Medium
Q5_K_M
5
19.4 GB
High
Q6_K
6
22.1 GB
High
Q8_0
8
28.9 GB
Very High
F16
16
55.4 GB
Maximum

Hardware-Kompatibilität

Eignungsschätzungen für alle Hardware

Rechner öffnen

Computing compatibility...

Speicheraufschlüsselung

Reference: RTX 2060 6GB

Weights7.2 GB
KV Cache1.0 GB
Runtime0.9 GB
Headroom0.6 GB

Häufig gestellte Fragen

FAQ — Ternary Bonsai 27B

Siehe auch