PrismML comprime o Qwen3.8 27B para 5,9 GB com pesos ternários, sob licença Apache 2.0
- O que aconteceu
- A PrismML lançou, a 17 de setembro de 2026, o Ternary Bonsai 2 27B, uma versão de 1,76 bits e 5,9 GB do Qwen3.8 27B, sob Apache 2.0.
- Porque é importante
- Um modelo aberto de 27B, com contexto de 262K e entrada de imagem, cabe agora numa GPU de consumo ou num Mac, com uma alegada retenção de 98,2 % da qualidade de precisão total.
- O que fazer
- Descarregue-o e teste-o nas suas próprias tarefas antes de substituir o modelo de precisão total.
Our verdict: conditional, so download it and test it. Ternary Bonsai 2 27B puts Qwen3.8-27B, a 262K-context text-and-image model, into 5.9GB under Apache 2.0. That is small enough for a single consumer GPU, a Mac, or even an iPhone or iPad. The catch: every number below comes from PrismML, and a 98.2% aggregate can hide a bigger drop on your task. If you run local models, this is worth an afternoon on your own workload before it replaces anything.
What happened
On 17 September 2026, PrismML released Ternary Bonsai 2 27B, weights on Hugging Face and GitHub under Apache 2.0. It stores Qwen3.8 27B as ternary weights (each -1, 0 or +1, with FP16 group-wise scaling), for 1.76 effective bits per weight. The result is 5.9GB, which PrismML says is more than 9x smaller than full precision.
PrismML's figures:
| Measure | Value |
|---|---|
| Aggregate score | 83.9, against 85.4 for full precision (98.2% retained) |
| Size | 5.9GB at 1.76 effective bits per weight |
| Context | 262K tokens |
| Input | Text and image |
| Throughput | Up to 143 tokens/s on RTX 5090, 46.8 tokens/s on M5 Max |
| Energy | 0.714 mWh per token on RTX 4090 |
| License | Apache 2.0 |
It runs on NVIDIA GPUs through CUDA, and on Macs, iPhones and iPads through MLX with PrismML's custom low-bit kernels.
Why it matters
At 5.9GB, a 27B open model with 262K context and image input fits the memory budget of a single consumer GPU or a Mac, and the Apache 2.0 license allows commercial use. If the 1.5-point aggregate gap holds on real work, local deployment of this model class gets far cheaper.
That "if" is why we hold it at conditional. Every figure is PrismML's own, and we have not tested it. An aggregate score averages many benchmarks, so a small average drop can mask a much larger drop on coding, vision or long-context tasks. The base model's own benchmarks are also vendor-reported, which is why Qwen3.8-27B sits at conditional too. You also need PrismML's kernels: a standard runtime will not do.
What changes for you
- If you run Qwen3.8-27B locally: run Ternary Bonsai 2 27B side by side on your own prompts. Swap only if the quality holds on the tasks you actually ship.
- If a 27B model was out of reach on your hardware: this is the one to try first. Pull the weights and run them through CUDA or MLX.
- If you need verified quality: wait for independent benchmarks. We have not found any yet.
O que fazer
- 1 Obtenha os pesos no Hugging Face ou no GitHub e execute-os com CUDA ou MLX.
- 2 Compare-o com o Qwen3.8-27B de precisão total na sua própria carga de trabalho. A retenção de 98,2 % é um valor agregado da PrismML.
Ferramentas e modelos afetados
Nunca mais precisas de te pôr a par
O resumo semanal: apenas mudanças de veredicto e ações urgentes. Sem enchimento.