On October 27, 2026, Liquid AI published the LFM3 family of small multimodal foundation models on the Hugging Face blog, together with a full recipe for turning the general-purpose base into domain-specific multimodal models: adapter architecture for vision and audio, a three-stage training pipeline, data mixing ratios, and on-device quantization and export scripts. The practical effect is that a small team with a single 24GB consumer GPU can now assemble image-text and speech models that
The October 27 post lays out the workflow in four steps.
The first is picking a base. LFM3 keeps the hybrid convolution-attention backbone of LFM2 and ships in 450M, 1.2B and 2.6B parameter tiers. Text, vision and audio share one tokenizer and decoder, with all modality differences absorbed by lightweight adapters that account for only 1.5% to 3% of total parameters.
The second is attaching modalities. The vision path uses a SigLIP-style ViT with a two-layer MLP projector; the audio path uses a log-mel front end with a 160ms frame shift. Liquid AI stresses that adapters train independently while the backbone stays frozen, so a single 24GB card is enough.
The third is staged training: modality alignment on large image-text and audio-caption corpora, instruction tuning with a 1:4 mix of general to domain data, then DPO to curb refusals and hallucination. In the published example,
Multimodal competition split visibly in 2026. One branch keeps scaling flagship models; the other squeezes vision and speech into 1B-class models, a path already taken by Gemma 3n, small Qwen3-VL variants and SmolVLM. Liquid AI paved that road with LFM2 in 2025 — a text base first, then LFM2-VL and LFM2-Audio, trading pure benchmark strength for edge throughput via a hybrid convolution architecture. The gap was never weights; it was know-how. Teams that downloaded the checkpoints still had to guess how to wire adapters, how to mix domain data, and how many points quantization would cost. LFM3's post fills that gap. Distribution matters too: Hugging Face lets the recipe, weights and example
Liu Gong's take: within six months, vertical multimodal models fine-tuned from 1B-class bases like LFM3 will ship as deliverable products in industrial inspection, equipment monitoring and report structuring — not as leaderboard entries. The reason is not raw capability but cycle time: publishing the recipe compresses delivery from roughly three months to three weeks.
From a technical standpoint, concentrating modality differences into adapters that hold just 1.5% to 3% of parameters, with the backbone frozen, is the right call for factory floors. The recurring constraint in AI visual inspection is not semantics but resolution: line cameras at 5 megapixels, fewer than 200 samples per defect class, and a hard rule that images never leave the plant. Training vision alignment and domain instruction tuning separately keeps catastrophic forgetting manageable and makes failures diagnosable. The caveat is that small models lose fine-grained localization long before they lose language ability; downscaling a 5MP frame to 384 pixels erases hairline scratches, so tiled inference remains mandatory and its engineering cost belongs in the budget.
On the business side, Liquid AI previously reported that LFM2-1.2B decodes on CPU at roughly twice the throughput of comparable Transformers. LFM3 keeps that hybrid backbone, so an edge box that served N camera streams should handle close to 2N — decisive in an inspection market priced per stream.
For the industry, the moat shifts from model tuning to data and delivery. Once the recipe is public, weights are commodity; annotation standards and process knowledge are not. Liu Gong predicts a wave of LFM3-based vertical vendors within a year selling judgment criteria rather than checkpoints.
- Parameter tiers: LFM3 ships at 450M, 1.2B and 2.6B, with multimodal adapters at 1.5%–3% of total parameters and a vision projector of about 8M parameters.
- Training cost: the published example uses 20,000 domain samples at LoRA rank 32, converging in about six hours on one A100 — under $30 at public cloud rates.
- Deployment footprint: the 1.2B multimodal model is roughly 0.7GB after 4-bit quantization and runs single-image inference on an 8GB phone.
- Speed comparison: LFM2-1.2B decoded on CPU at about twice the throughput of comparable Transformers, and LFM3 retains the same hybrid convolution architecture.
- Accuracy comparison: on the industrial part-defect and equipment-anomaly
For the industry, the near term (three months) brings a wave of community LFM3 fine-tunes and fast adapter support in llama.cpp and ONNX Runtime, shifting competition from benchmark rank to recipe usability. Within a year, on-device multimodality becomes the default for privacy-sensitive settings, cloud APIs lose ground wherever images cannot leave the building, and licensing plus SDK subscriptions become the main revenue line for small-model vendors.
For ordinary users, the near-term change is offline image Q&A and natural-language photo search on phones without a network round trip. Within a year the more meaningful gains land in accessibility and education: local models describing
Questions worth exploring:
- Where the division of labor between on-device small models and cloud flagships should sit, and whether LFM3's adapter scheme becomes a community standard.
- How open recipes reshape model vendors' business models: selling weights, selling SDKs, or selling delivered scenarios.
- In vertical settings such as industrial inspection, how to trade off the accuracy ceiling of small multimodal models against the engineering cost of tiled inference.
Further reading: 2026 Frontier LLM Comparison: A Practical Guide — a durable, evergreen reference. Bookmark it.
This item comes from Hugging Face Blog (published 2026-10-27 17:16:52). Wenfeng Radar provides a Chinese/English summary and commentary on overseas AI developments; the original content is owned by its author.
Daily aggregation of overseas AI developments with in-depth analysis. Bookmark us to never miss an important signal; back to homepage.