What GPU do I need to run prism-ml/Bonsai-27B-mlx-1bit?
1.7B parameters, published in F16. View on Hugging Face
Set up Bonsai-27B-mlx-1bitBonsai-27B-mlx-1bit is published by prism-ml on Hugging Face, with 1,919,048 downloads and 231 likes to date. It's a Qwen3_5ForConditionalGeneration model built for text-generation, published natively in F16. Its name carries the MLX tag.
VRAM required & cheapest live GPU fit
Required VRAM = weight size at each precision times a flat 1.2 overhead for activations and fragmentation. The KV-cache grows with context and is not in that factor; it is listed per model below where the architecture is published. Full formula and assumptions: methodology.
| Precision | Weight size | Required VRAM | Cheapest live fit | GPUs needed | Est. $/hr (full fit) |
|---|---|---|---|---|---|
| FP16 | 3.2 GB | 3.9 GB | V100 | 1 | $0.088/hr |
| FP8 (quantized) | 1.6 GB | 1.9 GB | RTX 4080 | 1 | $0.173/hr |
| INT4 (quantized) | 0.8 GB | 1.0 GB | RTX 3060 | 1 | $0.110/hr |
A GPU is only matched to a row if its hardware supports that precision, and the primary recommendation is always a single-GPU fit when one exists.
INT4 requires a quantized checkpoint actually published for this model, check its Hugging Face page before relying on this row.
Cheapest way to run Bonsai-27B-mlx-1bit at its published (F16) precision: 1× V100, at $0.088/hr per GPU ($0.088/hr total). Quantizing to FP8 or INT4 (rows above) can cost less, but requires a compatible quantized checkpoint to exist for this model.
Bonsai-27B-mlx-1bit: KV cache by context length
The KV cache is the memory the attention layers hold for every token of context, on top of the weights. It is not part of the flat 1.2x overhead in the table above, grows with context length and with every concurrent request, and is why a long-context deployment needs more VRAM than the table shows.
Hybrid: only the full-attention layers cache per token; the other layers keep a fixed-size state. Cached per token: 65,536 bytes at 16-bit.
| Context | KV cache, one sequence |
|---|---|
| 4K | 0.25 GB |
| 32K | 2.0 GB |
| 128K | 8.0 GB |
| 256K (model maximum) | 16.0 GB |
Computed from the layer, head and window counts in the model's own configuration (prism-ml/Bonsai-27B-mlx-1bit), at 16 bits per cached value. Each concurrent sequence needs its own cache, on top of the weights.
Bonsai-27B-mlx-1bit: common questions
How much VRAM does Bonsai-27B-mlx-1bit need?
3.9 GB at FP16, 1.9 GB at FP8 (quantized), 1.0 GB at INT4 (quantized). All of those sit under 6 GB, the smallest capacity this page reasons about, so VRAM is not what limits where this model runs. The 3.2 GB of weights plus inference overhead is the whole requirement.
How many copies of Bonsai-27B-mlx-1bit fit on one V100?
4, by VRAM alone. That card carries 16.0 GB and one copy needs 3.9 GB at FP16, on a live rate of $0.088/hr for the whole card. Throughput is not modelled here, so 4 copies is not 4 times the requests served.
Weight-to-VRAM math, the fit rules, and how live prices are normalized: full methodology.
More Qwen3.6 models
- Qwen3.6-27B-NVFP4 (21.2B, F8_E4M3)
- Qwen3.6-27B (27.8B, BF16)
- droplychee-1.0-27b (27.8B, BF16)
- EVE-27b-XENO-HAT-DeepSeek-V4-Flash (27.8B, BF16)
- Qwen3.6-27B-FP8 (27.8B, F8_E4M3)
Lineage
Based on Qwen3.6-27B.
Alternatives at this size
Other models for text-generation within about a third of Bonsai-27B-mlx-1bit's 1.7B parameters, from other model lines.
- Llama-3.2-1B-Instruct (1.2B, BF16)
- OLMo-2-0425-1B (1.5B, F32)
- LFM2.5-1.2B-Instruct (1.2B, BF16)
- SmolLM2-1.7B (1.7B, BF16)
- h2ovl-mississippi-2b (2.2B, BF16)
More on Bonsai-27B-mlx-1bit
Fits on an 8 GB GPU at FP16: every model that fits in 8 GB.
Related reading: V100 pricing and specs, How much VRAM you need for LLMs, Serving LLMs with vLLM, vLLM vs TensorRT-LLM vs SGLang, and Best GPU for LLM inference.
