Amoral LFM2.5-VL-1.6B
A refusal-removed ("abliterated") copy of LiquidAI/LFM2.5-VL-1.6B, produced with Heretic.
Why this exists
Every other model in this project targets a workstation or a rented GPU. This one targets a phone. At Q4_0 the weights are ~696 MB and the vision projector a further ~583 MB, and served at a 4,096-token context the whole thing peaks around 1.38 GB — small enough to run on a handset, and measured rather than estimated (see Provenance below).
What "amoral" means here
No refusal, not no doubt. This checkpoint is intended to engage with any request rather than decline. That is not the same as being right, and this project treats the difference as measurable: every published checkpoint carries a calibration measurement, and where ablation has made the model less well calibrated than the base it came from, the card says so plainly rather than omitting it.
A model that will discuss anything and is confidently wrong is less useful than one that refuses, because a refusal is visible and unearned certainty is not.
What this is not
- Not a large model. It is 1.6B. It is reliable on salient attributes — subject, colour, garment — and unreliable on spatial relations and fine detail. Treat its descriptions accordingly.
- Not fine-tuned. Abliteration only; no training on this project's caption corpus.
- Not a safety-tested artefact. It has had refusal behaviour deliberately removed.
Serving
Served from GGUF through llama.cpp rather than through Ollama, which is also what the on-device runtimes consume:
llama-mtmd-cli -m <model>.gguf --mmproj <mmproj>.gguf \
-c 4096 --image <path> -p "Describe this image."
The -c 4096 is not incidental: at the model's native 32,768-token context the same run peaks at 2.82 GB instead of 1.38 GB, for identical output.
License and attribution
Derived from LiquidAI/LFM2.5-VL-1.6B, licensed under the LFM Open License v1.0. That licence's copyright, patent, trademark and attribution notices are retained and apply to this derivative.
LFM Open License v1.0 is Apache-2.0-derived with one substantive change: commercial use rights end if your organisation's annual revenue reaches $10,000,000 USD. It carries no acceptable-use policy and no restriction on modifying safety behaviour.
Provenance
Machine-readable in run.json.
| trained from | LiquidAI/LFM2.5-VL-1.6B |
| method | heretic |
| pipeline commit | 7cf66f4cde3d8c5ffd618403a66fba851c1827a5 |
| run | lfm2-5-vl-1-6b-abliterate-20260920-201206 |
Measured
| capability | probe | metric | value | parent | |
|---|---|---|---|---|---|
| refusal | heretic-keywords-100 | refusal_keyword_count | 4 | 95 | held |
| vision | v2-ds4.1 held-out image 000058b4 | salient_attributes_correct | 1 | 1 | held |
Calibration
This checkpoint is less calibrated than the model it was trained from.
| metric | this checkpoint | parent | |
|---|---|---|---|
| ece | 0.2382 | 0.2123 | worse |
| brier | 0.2897 | 0.2837 | worse |
It will still engage with anything asked of it. What it has lost is accuracy about how sure to be — its expected calibration error moved from 0.2123 to 0.2382, the wrong way.
This is published rather than omitted because a card that hides a failed measurement is worse than one that never measured: it looks measured. See ADR-031.
On the instrument. Measured with the
confidenceelicitation, which is this architecture's own (spec 024): it answers 100/100 under that prompt and 20/40 under the full encoding, so the richer instrument measures it better on the rows it manages while discarding half of them. SENSITIVITY IS UNVALIDATED FOR THIS ARCHITECTURE. The same elicitation reports an ablation delta of +0.0007 for qwen3_5, where the encoding elicitation reports +0.0858 on the same checkpoints -- so it is demonstrably insensitive there. There is no second instrument this architecture can complete, so a PASS here means 'no regression detected by the only instrument that works', which is weaker than 'no regression'. Read it that way until spec 024 US2 supplies a claim set that admits a richer prompt.
Claim set: 77a2c609c86281ec.
Attribution
Derived from LiquidAI/LFM2.5-VL-1.6B, licensed
under LFM Open License v1.0. That licence's copyright and attribution
notices are retained here and apply to this derivative.
Apache-2.0-derived, with one substantive change: commercial use rights end if your organisation's annual revenue reaches $10,000,000 USD. It carries no acceptable-use policy and no restriction on modifying safety behaviour. §4(c) requires retaining all copyright, patent, trademark and attribution notices.