| --- |
| license: gpl-3.0 |
| base_model: HuggingFaceTB/SmolVLM2-500M-Video-Instruct |
| tags: |
| - gguf |
| - vision-language |
| - smolvlm2 |
| - arianna-method |
| - molequla |
| - notorch |
| --- |
| |
| # molequla — weights |
|
|
| Weights that molequla organisms use on the phone. Code lives at |
| [github.com/ariannamethod/molequla](https://github.com/ariannamethod/molequla); |
| the log book there (`MOLEQULALOG2.md`) carries the measurements quoted below. |
|
|
| ## ocelli/ — the eye |
|
|
| `ocelli` is the vision organ: SmolVLM2-500M with the Yent eye LoRA v2 merged |
| (SFT, 2026-05-29), converted to GGUF and run by a C inference path on |
| notorch. It gives an organism a sentence about a camera frame; that sentence |
| enters the colony as the fifth DNA source (`world`). |
|
|
| | file | quant | size | origin | |
| |---|---|---|---| |
| | `yent_eye_ours_q4_0.gguf` | Q4_0 | 232 MB | quantised on phone-1 with notorch `gguf_quantize` from the f16 merge | |
| | `yent_eye_ours_q6_k.gguf` | Q6_K | 418 MB | same | |
| | `yent_eye_ours_q8_0.gguf` | Q8_0 | 437 MB | same | |
| | `yent_eye_smolvlm2_lora_v2_q8_0.gguf` | Q8_0 | 437 MB | quantised on neo (llama.cpp) | |
| | `yent_eye_smolvlm2_lora_v2_mmproj_q8_0.gguf` | Q8_0 | 109 MB | vision tower + projector, required beside any decoder | |
| | `yent_eye_smolvlm2_lora_v2_mmproj_f16.gguf` | f16 | 199 MB | vision tower + projector | |
| |
| `SHA256SUMS.txt` in the folder lists every file; the f16 decoder merge |
| (820 MB) is not uploaded here. |
| |
| ### Measured on phone-1 (Galaxy A56, Exynos 1580, cores 4-7, OpenBLAS) |
| |
| One global 512×512 frame (`SMOLVLM_NOSPLIT=1`), 84-token prompt, from |
| `MOLEQULALOG2.md` (2026-09-13): |
|
|
| | decoder | gen tok/s | wall | peak RSS | |
| |---|---|---|---| |
| | `yent_eye_ours_q6_k` | 4.6 | 11.4 s | 951 MB | |
| | `yent_eye_smolvlm2_lora_v2_q8_0` | 6.4 | 13.5 s | 988 MB | |
| | f16 (not uploaded) | 1.4 | 31.2 s | 1602 MB | |
|
|
| Tiled prompts (13 frames, 878 tokens) cost 74-94 s per frame on this phone at |
| equal or worse content, so the eye runs on one global frame. The Q4_0 decoder |
| generated slower than Q8_0 on the tiled run (3.0 vs 5.0 tok/s) and dropped |
| objects from the description; Q6_K and Q8_0 reproduced the same sentence. |
|
|
| ## Coming |
|
|
| Other VLMs as alternative eyes for the same organ are planned; they will land in |
| sibling folders with the same measurement table. |
|
|
| ## ears/ — the hearing |
|
|
| `ears` is whisper on notorch: a C organ that loads OpenAI Whisper weights in |
| ggerganov's flat ggml format and transcribes 16 kHz mono wav with greedy decoding. |
| Its gate is parity with whisper.cpp: on phone-1 the log-mel is bit-identical, and |
| the transcripts of `jfk.wav` and of the same sentence re-recorded through the |
| phone microphone match token for token on both models below. Encoder outputs |
| differ by up to 4.9e-2 (mean 2.2e-4) because whisper.cpp keeps attention in f16. |
|
|
| | file | params | size | origin | |
| |---|---|---|---| |
| | `ggml-tiny.bin` | 39M, multilingual | 78 MB | OpenAI Whisper tiny, ggml conversion from whisper.cpp `models/download-ggml-model.sh` | |
| | `ggml-base.bin` | 74M, multilingual | 148 MB | OpenAI Whisper base, same | |
|
|
| Measured on phone-1, cores 4-7, four threads, `jfk.wav`: `ears` tiny 4.2 s wall, |
| 131 MB peak RSS. Whisper weights are MIT-licensed by OpenAI; `SHA256SUMS.txt` |
| in the folder lists both files. |
|
|
