Find the Best AI Model for Any Task from Hugging Face

A skill for finding the best Hugging Face model for a task by querying official HF benchmark leaderboards, enriching results with model size data, filtering for what fits the user's device, and returning a ranked comparison table. Step 1 parses the request for task (coding, math/reasoning, chat, OCR, RAG, speech recognition, image classification, multimodal, agents) and device (Apple Silicon unified memory, RTX VRAM, CPU-only, or no constraint) - if device isn't mentioned, size filtering is skipped entirely. Step 2 fetches the full list of official HF benchmark datasets via the datasets API filtered to benchmark:official and selects every dataset relevant to the task (not limited to 2-3 - use all that clearly apply). Step 3 fetches the top 15 ranked models from each selected benchmark's leaderboard, skipping and noting any leaderboard that 404s or 401s. Step 4 enriches the top 10-15 candidate model IDs with metadata via the models API or hf-cli, extracting parameter count from safetensors.total (converted to B) or parsed from the model name if safetensors metadata is absent, plus license from model-card tags. Step 5 filters and ranks: when a device is specified, models exceeding the fp16 parameter budget are removed unless they fit under Q4 quantization (roughly 4x the fp16 budget), in which case they're kept with a "needs Q4" note rather than silently dropped; models are then ranked by benchmark score and the top 5-8 kept, with proprietary models (GPT-4, Claude, Gemini) included but flagged "API only / not self-hostable" unless the user asked for local/open models only. The device-to-parameter-budget formula is fp16 max params (B) ~= memory (GB) / 2, Q4 max params (B) ~= memory (GB) x 2 - e.g. 16GB yields 8B fp16 / 32B Q4, 24GB VRAM yields 12B fp16 / 48B Q4. Step 6 outputs a markdown comparison table (model linked to its HF page, params, per-benchmark scores with "-" for unevaluated, license, and an "On device" column of Yes (fp16)/Q4 only/Too large/API only, starring the top pick), then asks whether the user wants to run the top model locally (with setup instructions) or via HF Jobs. Error handling covers a missing leaderboard (skip and note), a model missing metadata (parse size from its name), no benchmarks found for the task (fall back to hub_repo_search sorted by trendingScore), and all leaderboards failing (fall back to popularity-based search, disclosed as such).

Use it when a user asks about finding the best, top, or recommended model for a task, wants to know what AI model to use, or wants to compare models by benchmark score - triggering on phrasing like "best model for X" or "which model runs on my [device]".

Input is a task description and optionally a device/hardware constraint. Output is a ranked markdown comparison table of 5-8 models with benchmark scores, parameter count, license, and device-fit status, plus a follow-up offer to help run the top pick locally or via HF Jobs.

curl -s -H "Authorization: Bearer $(cat ~/.cache/huggingface/token)" "https://huggingface.co/api/datasets?filter=benchmark:official&limit=500"

Uses the Hugging Face Hub's datasets API (official benchmarks and their leaderboards), models API or hf-cli for model metadata, and hub_repo_search as a popularity-based fallback when benchmark data is unavailable.

Developers and researchers choosing which Hugging Face model to use for a task, especially those with a specific device to fit the model on - getting a benchmark-ranked comparison instead of manually cross-referencing multiple leaderboards and model cards.