# AI Model Release Research

> Investigate new AI model releases: weights, license, buzz.

- Skill: `wcpaka-lgtm/ai-model-release-research` (Agent Skill, multi-file: 2 files)
- Install (CLI): `npx skillmds@latest add wcpaka-lgtm/ai-model-release-research`
- Raw SKILL.md: https://api.skillmd.com/api/skills/wcpaka-lgtm/ai-model-release-research/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: AI & ML
- Author: wcpaka-lgtm (https://skillmd.com/u/wcpaka-lgtm)
- Updated: 2026-09-21
- Page: https://skillmd.com/skills/wcpaka-lgtm/ai-model-release-research

---


# AI Model Release Research

Investigate newly announced AI models: did the weights actually ship, what are the real specs, what does the license allow, and what do actual users report? The user asks this class of question often ("정식출시한거야?", "실사용 후기 알려줘"), sometimes as input to a blog post (see `blogger-publishing`). Exact API recipes and endpoints in `references/research-recipes.md`.

## Core discipline: separate vendor claims from verified facts
Vendor announcements mix positioning ("second only to X") with confirmed specs. Keep two buckets in every answer:
- **Confirmed**: weights downloadable, license text read, benchmarks independently reproduced, third-party leaderboard entries (Artificial Analysis, LMArena).
- **Claimed**: vendor benchmark tables (especially in-house named benchmarks), "open weights soon", parameter counts without active-parameter disclosure.
If the release is <48h old, say so explicitly — mature reviews don't exist yet; report what IS available (hardware questions, first impressions) instead of pretending.

## Verification ladder (teaser → actual release)
1. Announcement exists (X/blog) — but check whether it's a *preview* (API-only) or *weight release*.
2. Probe the official Hugging Face repo via API (recipes file): `createdAt`, file count, `likes`, license tag. A repo with 0 files or no LICENSE means weights haven't really shipped.
3. Read `LICENSE` raw — do NOT trust "open weights" framing. Community licenses may exclude territories or cap commercial revenue.
4. Check third-party surfaces: OpenRouter listing, Artificial Analysis, ComfyUI/engineering blogs for day-0 support.

## License checklist (CRITICAL for this user — they are in Korea)
Read the actual LICENSE file, specifically:
- **Excluded Territories** — some community licenses (e.g. MiniMax H3, Aug 2026) exclude **EU, UK, Republic of Korea, and USA**. If Korea is excluded, flag it prominently near the top of any summary/blog post; it changes the user's practical calculus entirely.
- Revenue cap for commercial use (e.g. $20M/yr threshold requiring separate authorization).
- Distillation / output-reuse restrictions, attribution requirements ("Powered by X" display).
- License NAME matters: "community license" ≠ Apache 2.0. Say which.

## Community sentiment — what works in this environment
Reddit comment content IS reachable via the **arctic-shift API** (`arctic-shift.photon-reddit.com/api/comments/search?link_id=<thread_id>&limit=100`) — full comments incl. scores, indexed within hours of day-0 posts. Get thread ids via web_search `site:reddit.com <topic>`, then bulk-fetch with curl inside execute_code, saving to temp files (stdout truncates ~20KB). Full recipe in the recipes file. Still dead: reddit `.json` endpoints, most redlib mirrors, pullpush for fresh posts.
Other working sources:
- **HN Algolia API**: `search_by_date?query=%22Model+Name%22&tags=comment` and `tags=story` — captures developer crowd reaction.
- **HF discussions API**: `/api/models/<org>/<repo>/discussions?limit=30` — surfaces license confusion, VRAM questions, regional complaints within hours of release.
- **GitHub search + PR API**: repos named after the model reveal ecosystem momentum; `api.github.com/repos/<org>/<repo>/pulls/<n>` (title/body/merged_at) verifies what actually shipped — e.g. whether day-0 support landed in core (Kijai's MiniMax-H3 support merged into ComfyUI core PR, NOT a separate wrapper repo). GitHub release-notes pages (`releases/tag/vX`) list every PR in the release.
- **Vendor engineering blogs** (e.g. ComfyUI blog) for concrete VRAM/quantization/hardware numbers — usually the only hard local-run data available early.
- For video models: mine arctic-shift comments with a hardware regex (`3060|3090|4090|vram|\d+\s?gb|minutes`) — real users post GPU + RAM + seconds-per-generation figures within 24h of weights dropping. Sort by score for signal.

## Hardware practicality (for local-run questions)
- Get exact file sizes via HF API `?blobs=true` (recipes file). Note bf16 vs int8/quantized variants.
- Watch for oversized text encoders (e.g. 32B+ vision-language encoders dominating memory).
- Report vendor-claimed minimum GPU with the optimization technique named (pruning, quant, offloading) — and label it as vendor-claimed.

## Pitfalls
- Param count ≠ serving cost for MoE: always find/disclose active parameters.
- Day-0 download counters on fresh HF repos are unreliable (often 0 while files exist).
- Name collisions: distinguish flagship names from similar-sounding small models (e.g. Qwen3.8 vs Qwen3-8B) — search engines conflate them.
- For blog output: hand findings to the `blogger-publishing` skill's tone rules (vendor claims labeled, honest "too early" caveats, Korean-user license flags).