Add GGUF model registry entries for Flare engine

[sauravpanda]

Summary

Add GGUF model entries to BrowserAI's model registry so users can load models via the Flare engine without specifying URLs.

Proposed models

Tier 1 — Small (instant load, <200MB)

Model ID	GGUF File	Size	Notes
smollm2-135m-flare	SmolLM2-135M-Instruct Q8_0	138MB	Fastest load, good for demos
smollm2-135m-flare-q4	SmolLM2-135M-Instruct Q4_K_M	~75MB	Smaller download

Tier 2 — Medium (good quality, <500MB)

Model ID	GGUF File	Size	Notes
smollm2-360m-flare	SmolLM2-360M-Instruct Q8_0	~350MB	Better quality
qwen2.5-0.5b-flare	Qwen2.5-0.5B-Instruct Q4_K_M	~350MB	Multilingual

Tier 3 — Large (best quality, ~1GB)

Model ID	GGUF File	Size	Notes
llama-3.2-1b-flare	Llama-3.2-1B-Instruct Q8_0	1.2GB	Best quality
llama-3.2-1b-flare-q4	Llama-3.2-1B-Instruct Q4_K_M	~600MB	Balanced

Registry format

{
    "smollm2-135m-flare": {
        engine: "flare",
        url: "https://huggingface.co/Qwen/SmolLM2-135M-Instruct-GGUF/resolve/main/smollm2-135m-instruct-q8_0.gguf",
        architecture: "llama",
        contextLength: 2048,
        quantization: "Q8_0",
        size: "138MB",
        features: ["chat", "instruction-following"]
    }
}

Advantage over MLC models

GGUF files are standard and available on HuggingFace without any conversion. Users can also load custom GGUF files not in the registry.

Tasks

• Find/verify GGUF download URLs on HuggingFace
• Add registry entries
• Test each model loads and generates correctly
• Add model cards to documentation

Depends on

• Implement FlareEngine adapter for BrowserAI engine interface #295 (FlareEngine adapter must be implemented)

Related

• Parent: Add Flare as a third inference engine backend #293

[sauravpanda]

Completed in PR #301 — flare-models.json includes 6 entries across 3 tiers: SmolLM2-135M (Q8_0, Q4_K_M), SmolLM2-360M (Q8_0), Qwen2.5-0.5B (Q4_K_M), Llama-3.2-1B (Q8_0, Q4_K_M).