Youssofal/Ternary-Bonsai-2-27B-MTPLX-Optimized-Speed: what it needs to run
Youssofal/Ternary-Bonsai-2-27B-MTPLX-Optimized-Speed is a merge of Qwen/Qwen3.8-27B. Its configuration matches Qwen3.8-27B's on every field that decides memory, so Qwen3.8-27B's requirements are this repository's.
Answered by
Geometrypublished data. Both configurations are published files, read at the revisions named below and compared field by field. Not a test that the model loads.Qwen3.8-27B (Qwen/Qwen3.8-27B) — every compared field is stated by both and equal.
Before you use those figures
- Its safetensors index counts 27,359,638,768 parameters against the base's 27,781,427,952 (1.5% apart). The compared geometry is the same, so the difference is in weights this comparison does not cover, and the base's file sizes are off by about that share for this repository.
Qwen3.8-27B: weight-file size by format
| Format | Size | Range | Basis |
|---|---|---|---|
| FP16 | 55.6 GB | 38.9 GB – 56.1 GB | size range |
| Q8_0 | 29.5 GB | 20.7 GB – 32.2 GB | size range |
| Q6_K | 22.8 GB | 16.0 GB – 26.0 GB | size range |
| Q5_K_M | 19.8 GB | 13.4 GB – 26.0 GB | size range |
| Q5_0 | 19.4 GB | 13.4 GB – 26.0 GB | size range |
| Q4_K_M | 17.1 GB | 10.9 GB – 26.0 GB | size range |
| Q4_0 | 16.2 GB | 10.9 GB – 26.0 GB | size range |
| Q3_K_M | 13.9 GB | 8.36 GB – 26.0 GB | size range |
| Q2_K | 11.0 GB | 6.38 GB – 26.0 GB | size range |
These are Qwen3.8-27B’s sizes. A model with the same geometry quantizes to the same size, to within any difference in parameter count noted above. Which devices hold each, and how fast they run it, is on Qwen3.8-27B’s page.
What was read, and where
- Youssofal/Ternary-Bonsai-2-27B-MTPLX-Optimized-Speed at revision 03bd60bb8275 is tagged by its publisher as a merge of Qwen/Qwen3.8-27B.
- Youssofal/Ternary-Bonsai-2-27B-MTPLX-Optimized-Speed/config.json at 03bd60bb8275 was compared with Qwen/Qwen3.8-27B/config.json at 1d4bf0f2ff60, the revision the catalogue pins: layers and their pattern, widths, head counts, experts, latent and recurrent dimensions, vocabulary, tied embeddings and any vision or audio tower.
Differences that do not change a memory answer
- architectures: null here, ["Qwen3_5ForConditionalGeneration"] in the base (a class name).
- Licence, as tagged
- apache-2.0
- Parameters in its safetensors index
- 27B
- Task, as tagged
- image-text-to-text
- Last changed on Hugging Face
- 2026-09-24
Publishing this model? A badge for its card
[](https://llmbottleneck.com/hf/Youssofal/Ternary-Bonsai-2-27B-MTPLX-Optimized-Speed)
It states the memory to run the model at Q4_K_M with 8,192 tokens of context — built on this repository’s own Q4_K_M file when it publishes one — and links back to this page, where the evidence is.
Resolve another repository
The same answer as JSON: GET /v1/resolve?repo=Youssofal/Ternary-Bonsai-2-27B-MTPLX-Optimized-Speed, with a free key from /keys.