Ternary Bonsai 2 9B — 显存、速度与本地部署
作者： Jakub Rusinowski · 最后更新：
8.95B 参数的语言模型加 0.46B 视觉塔，由 Qwen3.5-9B 以三值权重重建：32 个块中 8 个使用全注意力，24 个使用线性注意力。PTQ1_0 文件为 2.0 GB，PQ2_0 为 2.42 GB，Q2_0 为 2.56 GB——同一组权重的三种布局。262,144 token 上下文，支持图像输入和工具调用，另有单独文件内置 MTP 草稿头。思考模式即使面对简短提示也可能消耗 3,000 个以上 token，因此模型卡建议开启思考时至少预留 8,192 个输出 token。
Ternary Bonsai 2 9B 在 PTQ1_0 ternary 下约需 3 GB 显存——量化权重加框架开销，不含 KV 缓存。在 Apple Silicon 上，这部分来自统一内存。
按量化级别的显存与速度
计算基准：NVIDIA RTX 4090 (24 GB)。包含 8K 上下文的 KV 缓存，因此数值高于上方的主数字。
| 量化 | 显存 | 显存 | 速度（估算） | 适配 |
|---|---|---|---|---|
| PTQ1_0 ternary 1.78 bpw | 3.1 GB | ~198 tok/s | 可运行 | |
| PQ2_0 ternary 2.16 bpw | 3.5 GB | ~178 tok/s | 可运行 | |
| Q2_0 ternary 2.28 bpw | 3.6 GB | ~172 tok/s | 可运行 |
黑色标记 = NVIDIA RTX 4090 (24 GB) 上的可用显存。 估算来自内存带宽屋顶线模型，详见 方法说明页. Ternary Bonsai 2 9B 显存计算器 →
运行 Ternary Bonsai 2 9B
目录中能运行 Ternary Bonsai 2 9B 的最便宜 GPU 是 Intel Arc B570 (10 GB).
或在 Vast.ai 比较，低至 $0.35/小时 (typical low · varies)
作为亚马逊联盟成员，我们从符合条件的购买中获得收入。云 GPU 链接为推荐链接——我们可能获得佣金，您无需额外付费。
如何运行 Ternary Bonsai 2 9B
No Ollama library entry. The packed ternary formats need PrismML's llama.cpp fork (or upstream llama.cpp for the Q2_0 file only); on Apple devices use the MLX build with mlx-vlm 0.7.8 or newer.
规格
Corroborated — Two or more independent sources agree on these figures, but the model card itself was not retrieved. Treat the numbers as good rather than confirmed. 尚未确认： parameter count (model card only; the Hub's GGUF metadata reports 1.47B for this packed file).
- 参数量
- 8.95 Billion (+0.46B vision tower)
- 上下文窗口
- 262,144
- 架构
- Qwen3.5-9B hybrid attention (24 linear + 8 full), ternary weights
- 提供商
- PrismML
- 许可证
- Apache 2.0
- Availability
- Weights downloadable
- Practical deployment
- Edge device (phone, SBC)
- 规格量化
- PTQ1_0 ternary
- Vision projector
- +0.6 GB when you send images (not included above)
- 系统内存
- 8 GB
- 记录更新于
- 2026-10-15
Commercial use permitted. No usage restrictions beyond attribution.
限制与注意事项
- Ships only in PrismML's ternary packings; there is no Q4_K_M, Q8_0 or FP16 GGUF to choose instead.
- Upstream llama.cpp loads only the Q2_0 file; PTQ1_0, PQ2_0 and the drafters need the PrismML fork. The released mlx-lm (0.32.0) cannot load the MLX build yet; mlx-vlm 0.7.8 can.
- The vision projector is a separate file (0.62 GB at Q8_0).
质量与使用场景
评分由模型作者或独立评测方发布——衡量质量而非吞吐量，并非我们实测。
我的 GPU 能运行 Ternary Bonsai 2 9B 吗？
Ternary Bonsai 2 的其他尺寸
Ternary Bonsai 2 9B — 常见问题
How much VRAM does Ternary Bonsai 2 9B need?
About 3 GB at PTQ1_0 ternary — quantized weights plus framework overhead, before any KV cache. The cache grows with context length and is added on top; the table above folds it in. Apple Silicon counts unified memory toward the same figure.
Does Ternary Bonsai 2 9B run on an RTX 4090 (24 GB)?
Yes. Ternary Bonsai 2 9B needs about 3 GB at PTQ1_0 ternary, inside a 24 GB card, at an estimated 198 tokens/sec.
How do I run Ternary Bonsai 2 9B locally?
No Ollama library entry. The packed ternary formats need PrismML's llama.cpp fork (or upstream llama.cpp for the Q2_0 file only); on Apple devices use the MLX build with mlx-vlm 0.7.8 or newer. Running the published tag would send your prompts to a hosted GPU rather than your own machine.
What other sizes does Ternary Bonsai 2 come in?
Ternary Bonsai 2 2B (1 GB), Ternary Bonsai 2 4B (2 GB), Ternary Bonsai 2 9B (3 GB), Ternary Bonsai 2 27B (7 GB). Every size shares the family's training and licence; the larger ones score higher and need proportionally more memory.