--- title: Global LLM Download Leaderboard emoji: 🌍 colorFrom: yellow colorTo: indigo sdk: static pinned: false license: cc-by-4.0 short_description: HF text-generation, trailing 30-day downloads by country tags: - leaderboard - llm - downloads - open-models - korea - analytics --- # Global LLM Download Leaderboard **Which open language models people actually download β€” and where those models came from.** Ranking axis is **trailing 30-day downloads** on Hugging Face `text-generation`, not cumulative totals. Cumulative figures reward age; the 30-day window shows what is being used now. Both are shown for every row. - **Global Top 300** Β· **Korea Top 30** (Korean organisations) - Country shares, parameter size, quantization type, model type, release year, top organisations ## How country is attributed **By model family β€” who trained it β€” not by the uploading organisation.** A third-party quantization of a model counts toward the family that trained the original. We resolve the family from the model's **architecture** (`model_type`) rather than its name, because names can be changed and architectures cannot. Repositories that are quantized or converted variants inherit the attribution of the model they were derived from. Entries we cannot attribute from that evidence are left as **Other**. We do not guess. The unresolved share is printed in the footnote on the page. ## What is excluded CI test fixtures (`internal-testing`, `tiny-random-*`, and similar) are removed. They accumulate large download counts from automated test suites and are not models anyone uses. Embedding and reranker repositories are also removed. The platform tags many of them `text-generation`, but they do not generate text, and at their download volume they would displace the models this table exists to rank. ## Data and limits Measured from the Hugging Face API on the date shown on the page. This is a snapshot, not a live feed; the measurement date is printed at the top and in every export. Download counts are what the platform reports. They include automated and mirrored traffic that we cannot separate, so they measure **distribution**, not users. Read them that way. ## License Page and methodology released under **CC BY 4.0**. Download figures are property of their source platform. Built by [VIDRAFT](https://vidraft.net). --- # κΈ€λ‘œλ²Œ LLM λ‹€μš΄λ‘œλ“œ λ¦¬λ”λ³΄λ“œ **μ–΄λ–€ 곡개 μ–Έμ–΄λͺ¨λΈμ΄ μ‹€μ œλ‘œ λ‚΄λ €λ°›μ•„μ§€κ³  있으며, κ·Έ λͺ¨λΈμ€ μ–΄λ””μ—μ„œ μ™”λŠ”κ°€.** μˆœμœ„ 기쀀은 ν—ˆκΉ…νŽ˜μ΄μŠ€ `text-generation`의 **졜근 30일 λ‹€μš΄λ‘œλ“œ**μž…λ‹ˆλ‹€. λˆ„μ μ΄ μ•„λ‹™λ‹ˆλ‹€. λˆ„μ μ€ 였래된 λͺ¨λΈμ— μœ λ¦¬ν•˜κ³ , 30일 창은 **μ§€κΈˆ μ“°μ΄λŠ” 것**을 λ³΄μ—¬μ€λ‹ˆλ‹€. ν‘œμ—λŠ” λ‘˜ λ‹€ μ‹€λ¦½λ‹ˆλ‹€. - **κΈ€λ‘œλ²Œ Top 300** Β· **ν•œκ΅­ Top 30**(κ΅­λ‚΄ 쑰직) - ꡭ가별 점유 Β· νŒŒλΌλ―Έν„° 규λͺ¨ Β· μ–‘μžν™” μ’…λ₯˜ Β· λͺ¨λΈ μœ ν˜• Β· 곡개 연도 Β· 쑰직별 μƒμœ„ ## κ΅­κ°€ 귀속 κΈ°μ€€ **μ—…λ‘œλ” 쑰직이 μ•„λ‹ˆλΌ γ€Œλͺ¨λΈ 계열 β€” λˆ„κ°€ ν•™μŠ΅ν–ˆλŠ”κ°€γ€** κΈ°μ€€μž…λ‹ˆλ‹€. 제3μžκ°€ 올린 μ–‘μžν™”νŒμ€ **원본을 ν•™μŠ΅ν•œ 계열**둜 κ·€μ†ν•©λ‹ˆλ‹€. 계열은 이름이 μ•„λ‹ˆλΌ λͺ¨λΈμ˜ **μ•„ν‚€ν…μ²˜(`model_type`)** 둜 νŒμ •ν•©λ‹ˆλ‹€. 이름은 λ°”κΏ€ 수 μžˆμ§€λ§Œ μ•„ν‚€ν…μ²˜λŠ” λ°”κΏ€ 수 μ—†κΈ° λ•Œλ¬Έμž…λ‹ˆλ‹€. μ–‘μžν™”Β·λ³€ν™˜νŒμ€ κ·Έ μ›λ³Έμ˜ 귀속을 κ·ΈλŒ€λ‘œ λ¬Όλ €λ°›μŠ΅λ‹ˆλ‹€. 이 근거둜 νŒμ •λ˜μ§€ μ•ŠλŠ” ν•­λͺ©μ€ **'기타'둜 λ‘‘λ‹ˆλ‹€. μΆ”μΈ‘μœΌλ‘œ μ±„μš°μ§€ μ•ŠμŠ΅λ‹ˆλ‹€.** λ―Έν•΄κ²° λΉ„μœ¨μ€ νŽ˜μ΄μ§€ 각주에 μ‹€μΈ‘κ°’μœΌλ‘œ ν‘œκΈ°ν•©λ‹ˆλ‹€. ## μ œμ™Έ λŒ€μƒ CI ν…ŒμŠ€νŠΈ ν”½μŠ€μ²˜(`internal-testing`, `tiny-random-*` λ“±)λŠ” μ œμ™Έν•©λ‹ˆλ‹€. μžλ™ν™”λœ μ‹œν—˜μ—μ„œ λ‹€μš΄λ‘œλ“œκ°€ 크게 μŒ“μ΄μ§€λ§Œ μ‹€μ œλ‘œ μ“°λŠ” λͺ¨λΈμ΄ μ•„λ‹™λ‹ˆλ‹€. **μž„λ² λ”©Β·λ¦¬λž­μ»€ μ €μž₯μ†Œλ„ μ œμ™Έν•©λ‹ˆλ‹€.** ν”Œλž«νΌμ΄ μ΄λ“€μ—κ²Œλ„ `text-generation` νƒœκ·Έλ₯Ό λΆ™μ΄μ§€λ§Œ ν…μŠ€νŠΈλ₯Ό μƒμ„±ν•˜μ§€ μ•ŠμœΌλ©°, λ‹€μš΄λ‘œλ“œ 규λͺ¨κ°€ μ»€μ„œ κ·ΈλŒ€λ‘œ 두면 이 ν‘œκ°€ μ„Έλ €λŠ” λͺ¨λΈλ“€μ„ λ°€μ–΄λƒ…λ‹ˆλ‹€. ## 데이터와 ν•œκ³„ νŽ˜μ΄μ§€μ— ν‘œκΈ°λœ λ‚ μ§œμ— ν—ˆκΉ…νŽ˜μ΄μŠ€ API둜 μ‹€μΈ‘ν•œ **μŠ€λƒ…μˆ**μž…λ‹ˆλ‹€. μ‹€μ‹œκ°„ 갱신이 μ•„λ‹ˆλ©°, μΈ‘μ • λ‚ μ§œλ₯Ό νŽ˜μ΄μ§€ 상단에 항상 λ³‘κΈ°ν•©λ‹ˆλ‹€. λ‹€μš΄λ‘œλ“œ μˆ˜λŠ” ν”Œλž«νΌμ΄ λ³΄κ³ ν•˜λŠ” κ°’μž…λ‹ˆλ‹€. μžλ™ν™” νŠΈλž˜ν”½κ³Ό 미러 νŠΈλž˜ν”½μ΄ μ„žμ—¬ 있고 저희가 뢄리할 수 μ—†μœΌλ―€λ‘œ, 이 μˆ˜μΉ˜λŠ” **이용자 μˆ˜κ°€ μ•„λ‹ˆλΌ λ°°ν¬λŸ‰**을 λ‚˜νƒ€λƒ…λ‹ˆλ‹€. κ·Έλ ‡κ²Œ 읽어 μ£Όμ‹­μ‹œμ˜€. ## λΌμ΄μ„ μŠ€ νŽ˜μ΄μ§€μ™€ 방법둠은 **CC BY 4.0**. λ‹€μš΄λ‘œλ“œ μˆ˜μΉ˜λŠ” 좜처 ν”Œλž«νΌμ— κ·€μ†λ©λ‹ˆλ‹€. μ œμž‘ [λΉ„λ“œλž˜ν”„νŠΈ](https://vidraft.net).