Anima Light (GGUF)
A 1B companion for 4 GB phones: as accurate as its base on facts and memory, in a fifth of the words, fast to read prompts on Arm, and half as likely as its predecessor to make up a personal detail it was never given.
Anima Light is a LoRA fine-tune of Gemma 3 1B-it, merged and quantized to GGUF (0.72 GB). It is the free model in AI-Diary, a private journaling app that runs entirely on the phone, and the sibling of Anima Deep.
Files
| file | size | notes |
|---|---|---|
anima-light-v12-Q4_0.gguf | 0.72 GB | current / official production (2026-10-05). Q4_0 |
anima-light-v14.3-Q4_0.gguf | 0.72 GB | research candidate (Terse-DPO). Kept for reproducibility; not for app use (see note below) |
anima-light-v11-Q4_0.gguf | 0.72 GB | previous version, kept for app builds that ask for it by name |
anima-light-v8-Q4_K_M.gguf | 0.81 GB | older version, kept for older app builds |
sha256 of v12: 07e26d8cf0f25dd4fdbc42150efc355f930b2c3acfb5bd5e401f8f2da1674749
sha256 of v14.3: ce7f85c527db494a29fef7f8bac47fe1744c050d0c503e4aed2ebe86961c5664
sha256 of v11: d0f802541b20a7abaf668609c7af61539554f10201de22d80a27c8cf8fb514a4
Note: Why v12 remains the official production model over v14.3
A candidate fine-tune using DPO (anima-light-v14.3-Q4_0.gguf) was evaluated against v12 across our test battery:
- Factual grounding & false premises: v12 significantly outperforms v14.3 in avoiding hallucinations. On false premises, v12 fabricates answers in 45/72 cases vs 66/72 for v14.3 (v14.3 degraded below the base Gemma 3 1B model at 58/72).
- Answer degeneration: v14.3 re-introduced sub-sentence and one-word replies (47 of 780 answers were 1-2 words such as "Luis", "24", "Sé", often incorrect), whereas v12 has only 3.
- Language consistency: v14.3 showed bilingual leakage (answering Spanish questions entirely in English in 9/780 cases).
- Evaluation gates: v14.3 passed only 1 of 7 validation gates (memory). On our multi-layer capability index, v12 scores 75% overall versus 66% for v14.3 (base model: 67%).
Consequently, Anima Light v12 remains the official and active model deployed in AI-Diary.
Why Q4_0. Gemma 3 1B's tensor shapes don't fit the K-quant super-blocks, so a "Q4_K_M" file of this model is really ~40 % Q5_0 and ~40 % Q8_0. Q4_0 is real 4-bit on the transformer weights, and llama.cpp repacks it for Arm's dot-product instructions at load time. On a Pixel 7 Pro, the same model read the prompt 2.5-3x faster in Q4_0 than in Q4_K_M (41-66 → 109-169 tokens/s).
Usage
Gemma 3 has no system role: put the instructions at the start of the first user turn, as the app does.
llama-cli -m anima-light-v12-Q4_0.gguf
The model takes its name from the instructions it receives; no name or origin was trained in.
Evaluation
Our own suites, run on this GGUF with the app's real prompt (memory, history and language anchors filled in) and the app's sampling (temperature 0.2 in the short mode, 0.62 in the full mode). They are not standard public benchmarks.
| Gemma 3 1B-it | Anima Light v8 | Anima Light v11 | Anima Light v12 | |
|---|---|---|---|---|
| Basic facts, full mode / short mode (of 60) | 53 / 48 | 39 / 40 | 49 / 52 | 54 / 57 |
| Numeric facts in Spanish / English (of 24) | 17 / 22 | 6 / 12 | 22 / 23 | 18 / 23 |
| Uses a fact given earlier (of 48) | — | 17 | 46 | 46 |
| …when the previous turn was an "I don't have that" (of 48) | — | 22 | 39 | 48 |
| States a personal detail it was never given (of 120, lower is better) | 72 | — | 80 | 37 |
| Goes along with a false premise or a thing that doesn't exist (of 72, lower is better) | 58 | — | 69 | 45 |
| Identity (6) · foreign identities never trained ("Are you Grok?"…, 10) | 5 · 8 | 5 · 8 | 5 · 10 | 5 · 10 |
| Helpfulness when asked for help (of 50) | 44 | — | 31 | 32 |
| Median words, chat in short / full mode | 92 / 237 | 17 / 16 | 16 / 42 | 17 / 56 |
| One- or two-word answers | 5 | 105 | 0 | 1 |
| Prompt scaffolding leaked into the answer | 15 | 0 | 0 | 0 |
| Sentences mixing informal "tú" with impersonal "-se" (of 50) | 9 | 2 | 3 | 4 |
The two "lower is better" rows are 32 questions the model cannot answer (20 personal details nobody gave it, 12 false premises or invented things such as a novel that doesn't exist), in both modes, three sampling seeds. They are scored by pattern matching and read by hand in part; differences of a few points mean nothing, the halving does.
The public half of our pressure and identity suites is published as Anima Bench (ES), with a runner and scorer you can use on any model. There v12 scores 79/126 and v11 76, close to the base (73) and below v8 (91): v8 was trained to hold its position under pressure and v11 and v12 are not (see Limitations).
How to read this: at 1B, seeds of the same recipe differ by several points on every row (on "states a personal detail it was never given", the three seeds behind v12 scored 26, 52 and 62). v11 and v12 are each the average (a "soup") of three seeds rather than the best single seed, to ship the recipe and not the luck of one run.
Limitations
- Helpfulness is below the base model (31 vs 44 of 50): asked for help, it answers in a few sentences where the base writes a list. Two data rounds aimed at this did not beat seed noise.
- It still invents when it doesn't know, about half as often as v11. Asked for a personal detail it was never given, v12 states something in 37 of 120 cases (v11: 80, the base: 72), and it goes along with a false premise in 45 of 72 (v11: 69, the base: 58). It will still tell you which ocean borders a landlocked country, or name an author for a novel that doesn't exist. Some of its refusals are muddled ("you haven't told me yet; I have it saved for you"). AI-Diary also handles this outside the model: when the question asks for a stored personal fact that is nowhere in the prompt, the app answers "I don't have that saved" without generating.
- Its Spanish is uneven: grammatical most of the time, with occasional twisted sentences or invented words.
- Rarely, a stray token from another script (Bengali, Cyrillic) can appear mid-word; about 1 in 800 answers for v11. AI-Diary removes it with a sampling grammar that rejects non-Latin scripts unless the conversation needs them.
- It gives in under pressure, more than the base model. v8 was trained to hold its position and that distorted too much else at this size. v12 has a dozen such examples and they changed nothing measurable: on our 30-case pressure battery with the app's prompt, three sampling seeds each, v12 holds its position in 2-4 cases, v11 in 3-5 and the base in 6-11. If someone insists on something false, it tends to agree.
- Not a therapist. An app using it should route crisis situations to human help.
Training
- Rebuilt from the base for v10-v12 after v8: LoRA rank 8, scale 5 (v8 used 20), all 26 layers, lr 2e-5, 4 epochs, batch 8 (MLX, on the 4-bit base).
- v11: 214 hand-written SFT examples in Spanish and English, built with the app's real prompts: short and full-mode conversation, basic facts, using memory and history, identity, and long answers when asked. Always in informal "tú".
- v12: those 214 plus 148 about not inventing (362 in all): the same kind of personal question with the fact in the prompt (it uses it) and without it (it says it doesn't have it); false premises it corrects and invented things it says it doesn't know, next to real facts of the same shape so it doesn't start denying what it knows; questions whose fact is present right after a turn where the assistant said "I don't have that"; and twelve about not flattering.
- Each version averages the LoRA deltas of three seeds, merged into the original bf16 weights, converted with llama.cpp and quantized to Q4_0.
What we learned at 1B: the sentence is the reasoning. Pushing the model to answer below one full sentence breaks its facts (v8 got basic facts wrong for that reason), and a strong LoRA scale did the rest. The model also multiplies any phrasing pattern in its data: twelve impersonal openings ("Equivocarse en el trabajo duele") were enough to make it mix "tú" and "se" in its advice, and rewriting those twelve fixed it. The same lever works on purpose: v12's first round of "say you don't have it" examples made the model say it right after any earlier refusal in the conversation, even with the fact in front of it (28-38 of 48), and sixteen examples of exactly that situation brought it to 48 of 48.
Contact
LinkedIn · christeck.com · questions and ideas in the Community tab. Open to collaborations and consulting on on-device and small-model fine-tuning.
License
Gemma Terms of Use. Gemma is provided under and subject to the Gemma Terms of Use found at ai.google.dev/gemma/terms. Use is also subject to the Gemma Prohibited Use Policy. This is a modified version of Gemma 3 1B-it (see NOTICE).
Built by Christeck with Claude (Anthropic), who co-designed the data, evaluations and experiments.
Resumen en español
Anima Light es un ajuste fino de Gemma 3 1B para teléfonos de 4 GB, rehecho desde la base: acierta los datos básicos y usa la memoria como el modelo original, contesta en una quinta parte de las palabras, habla de tú y, en Q4_0, lee el prompt 2.5-3 veces más rápido en un teléfono Arm. La versión 12 inventa la mitad que la 11 cuando le preguntan un dato personal que nadie le dio (37 de 120 contra 80) y un tercio menos ante una premisa falsa (45 de 72 contra 69). Sus límites: todavía inventa, ayuda menos que la base cuando se le pide ayuda, su español a veces se tuerce y cede cuando le insisten en algo falso, más que el modelo original.
Nota de producción: Anima Light v12 es la versión oficial y activa para AI-Diary. El candidato experimental v14.3 se conserva únicamente para fines de investigación y trazabilidad, ya que en la auditoría exhaustiva v12 demostró superioridad fáctica (75% de índice general vs 66%), mejor rechazo de premisas falsas y ausencia de respuestas monosilábicas degradadas.