Model Catalog

The standard lightweight sentence embedding model for browser use. 384-dim output, excellent speed-to-quality ratio. Best pick for local semantic search.

BAAI's BGE-small-en-v1.5 — competitive quality vs MiniLM at nearly the same size. 384-dim, strong on MTEB benchmarks for a sub-30 MB model.

BAAI BGE-base-en-v1.5 — higher quality than the small variant. 768-dim, noticeably better retrieval on long documents. Worth the extra 85 MB.

Nomic AI's open embedding model. 768-dim with 8 192-token context window — significantly longer than MiniLM. Good for embedding long documents.

Apple's MobileViT-small — a hybrid CNN + Vision Transformer at 22 MB. Excellent for real-time browser image classification. Top pick for mobile and low-memory devices.

Google's Vision Transformer base model. Higher accuracy than MobileViT at 10× the size. Suitable when quality matters more than speed.

A 6M-parameter detection transformer based on DeiT-Tiny. At 12 MB it is the lightest serious object detector available for browser inference as of May 2026.

Facebook's DETR with ResNet-50 backbone. Higher accuracy than YOLOS-Tiny but 14× larger. Expect noticeable latency in browser — prefer YOLOS-Tiny for real-time use.

OpenAI Whisper Tiny, English-only. At 75 MB it is the smallest Whisper variant — best-effort transcription speed in browsers. Expect 5–15× slower than real-time without GPU.

OpenAI Whisper Base — 74M parameters, supports 99 languages. Better accuracy than Tiny at ~2× the size. Slow in browsers without WebGPU acceleration.

HuggingFace's SmolLM2 at 135M parameters — the smallest instruction-tuned LLM with usable output quality. Available in q4f16 via Transformers.js v3. Expect 80–180 tokens/min on GPU.

SmolLM2 at 360M parameters — noticeably better coherence than the 135M variant. q4f16 via Transformers.js v3. Expect 40–100 tokens/min on GPU.

Alibaba's Qwen2.5 at 0.5B parameters. Strong reasoning and instruction-following for its size. Available via onnx-community in q4f16. A step up from SmolLM2 in output quality.

Google's Gemma 3 at 1B parameters — released March 2025. Excellent instruction-following for a 1B model. Available in ONNX web format with q4f16. Requires WebGPU for practical speed.

Meta's Llama 3.2 1B — released September 2024, widely supported in ONNX format. Practical upper limit for browser LLM inference as of May 2026. Requires a capable GPU and 6+ GB RAM.

DeepSeek's R1 reasoning model distilled into Qwen 1.5B — released January 2025. Includes chain-of-thought reasoning. Heavy for browsers: ~1 GB download, WebGPU required, 8+ GB RAM recommended.

Meta's Segment Anything Model (SAM) with ViT-Base backbone. Click-based image segmentation — impressive quality, but 375 MB download makes it slow to load on first use.