VertexAGI/prism-caption-2-micro

Prism Caption 2 Micro

Prism Caption 2 Micro is a chat-titling model — given the first user message of a conversation, it generates a short, specific, correctly-formatted title (4-6 words, title case, naming the actual subject). Fine-tuned via LoRA on LiquidAI's LFM2-700M, switching base models from the Qwen3-0.6B lineage used in 1 Mini/1.5 Micro.

Part of the Prism family of small, single-purpose models.

Why LFM2-700M

LFM2-700M was chosen over the previous Qwen3-0.6B base on Liquid AI's own published evaluations:

Model	Params	MMLU	IFEval
Qwen3-0.6B	~600M	44.93	64.24
LFM2-700M	742M	49.9	72.23

LFM2-700M beats Qwen3-0.6B on both general knowledge/reasoning and instruction-following, and Liquid AI's architecture is purpose-built for edge/local inference — reported at roughly 2x Qwen3's CPU decode/prefill throughput at a comparable parameter count. For a small, single-purpose model meant to run cheaply and locally, that combination of higher instruction-following accuracy and faster inference made it a clear upgrade over the previous base.

Model Details

Base model	LiquidAI/LFM2-700M
Fine-tuning base checkpoint	mlx-community/LFM2-700M-4bit
Architecture	LFM2 -- hybrid short-convolution / attention (16 blocks total, 6 full-attention, 10 short-conv)
Fine-tuning method	LoRA (rank 8, scale 20.0, 16 layers)
Fine-tuning framework	MLX / mlx-lm, on Apple Silicon
License	LFM Open License v1.0

Training Data

Fine-tuned on 10,000 examples (9,000 train / 1,000 validation) of synthetic chat-titling pairs, distilled across a mix of four teacher models as their individual free-tier availability shifted over the course of generation:

Teacher	Examples	Share
openai/gpt-oss-20b (NIM)	5,339	53.4%
nvidia/nemotron-3.5-lightning-30b-a3b (NIM)	3,201	32.0%
poolside/laguna-s-2.1:free (OpenRouter)	1,027	10.3%
openai/gpt-oss-120b (NIM)	433	4.3%

1,207 unique topics, 6,601 unique generated titles. The teacher-cycling approach (alternating every ~2 minutes between models) was adopted mid-run after single-teacher generation kept stalling on free-tier rate limits -- spreading load across several capable teachers kept throughput high without any provider getting hammered continuously.

Training Procedure

• Method: LoRA fine-tuning (rank 8, scale 20.0, dropout 0.0, 16 layers -- full model depth)
• Optimizer: Adam, learning rate 1e-5
• Sequence length: 256 tokens
• Training steps: 6,000 iterations, validation every 200 steps
• Final validation loss: 0.196 (down from 8.054 at initialization) -- the final checkpoint had the best validation loss of the run, no earlier checkpoint needed
• Throughput: ~1.22 it/sec, ~537 tokens/sec, peak memory 1.6GB -- fast and light even on a memory-capped consumer machine

Evaluation

Compared against base LFM2-700M and the previous-generation Prism Caption 1.5 (Qwen3-0.6B) on a hand-authored held-out topic set with zero overlap against the training bank. Run three times at increasing scale to confirm the result wasn't a small-sample artifact:

Held-out size	Base LFM2-700M (issues / relevant / avg words)	Prism Caption 1.5	Prism Caption 2
24 topics	8/24, 18/24, 7.5w	1/24, 21/24, 3.3w	0/24, 24/24, 4.7w
145 topics	35/145, 107/145, 5.8w	9/145, 129/145, 3.9w	0/145, 145/145, 5.0w
275 topics	69/275, 205/275, 5.9w	13/275, 249/275, 4.0w	0/275, 273/275, 5.0w

"Issues" = formatting problems (too long/short/terse, leaked preamble, trailing punctuation, multiline). "Relevant" = title shares a non-stopword with the source topic.

Prism Caption 2 produced zero formatting issues across all 275 held-out topics in the final run, with 99.3% relevance -- the strongest and most consistent result of any Prism Caption generation. It also directly fixes 1.5's known failure mode: 1.5 averaged 3.3-4.0 words and occasionally over-compressed into vague titles ("Shoulder Sleep Matters" for a mattress question), while v2 lands at a natural ~5-word average and consistently names the actual subject ("Choosing Mattress for Side Sleepers").

Formats available

This repo includes both:

Format	File	Notes
MLX (4-bit)	model.safetensors + config	For Apple Silicon via mlx-lm
GGUF (Q4_K_M)	prism_caption_2_micro_Q4_K_M.gguf	For llama.cpp and compatible runtimes (LM Studio, Ollama, etc.)

Usage -- MLX

from mlx_lm import load, generate

model, tokenizer = load("VertexAGI/prism-caption-2-micro")

messages = [{"role": "system", "content": (
    "You name chat conversations. Given the user's first message, reply with ONLY a short, "
    "specific chat title (4-6 words, title case, no quotes, no punctuation at the end, no "
    "preamble). The title MUST name the main subject of the message -- do not over-abbreviate "
    "into something vague. Nothing else -- just the title."
)}, {"role": "user", "content": "Any advice on how to fix a leaking kitchen faucet?"}]
text = tokenizer.apply_chat_template(messages, add_generation_prompt=True, tokenize=False)

response = generate(model, tokenizer, prompt=text, max_tokens=24)
print(response)

Usage -- GGUF (llama.cpp)

llama-cli -hf VertexAGI/prism-caption-2-micro -m prism_caption_2_micro_Q4_K_M.gguf \
  -sys "You name chat conversations. Given the user's first message, reply with ONLY a short, specific chat title (4-6 words, title case, no quotes, no punctuation at the end, no preamble). The title MUST name the main subject of the message -- do not over-abbreviate into something vague. Nothing else -- just the title." \
  -p "Any advice on how to fix a leaking kitchen faucet?"

Limitations

Trained on a synthetic dataset distilled from a shifting mix of teacher models rather than a single consistent one; some stylistic inconsistency between examples from different teachers may be present. Titling behavior is only validated on English, conversational, everyday-topic inputs -- highly technical or non-English inputs are untested.

License

LFM Open License v1.0, inherited from the LFM2-700M base model. Free for research/non-commercial use and for commercial use under $10M annual revenue.