TrendHub Logo
TrendHub

Models Intelligence

Hugging Face and model radar

This workspace now blends Hugging Face launches with frontier benchmark context, open-weight watchlists and official model specs. GitHub repository scanning stays on the main Trend Radar page.

Performance Atlas

Frontier rankings, open-weight leaders and official model specs

This layer combines the live OpenAI model catalog, benchmark watchlists and vendor documentation so the model board is not limited to Hugging Face launch activity alone.

Synced Jul 30, 2026 3-source stack

Frontier Models

Closed-model performance watch

Provider APIs + benchmark watchlists + vendor docs

OpenAI

GPT-5.6 Luna

Frontier

OpenAI lane is now sourced from the OpenAI Models API Auto-Sync (2026-07-30). The site promotes the newest available GPT family model first and keeps the previous static GPT-5.4 entry only as historical fallback context.

Catalog source

Live API

Internal rank

ELO 2002

Context

1M

Model family

GPT

OpenAI modelsOpenAI model API

Anthropic

Claude Opus 5

Frontier

Anthropic lane now tracks the OpenRouter Catalog Auto-Sync (2026-07-30). The newest Claude family flagship is promoted automatically as soon as it appears in the synced catalog.

Catalog source

Live API

Internal rank

ELO 1870

Context

1M

Max output

64k+

Claude models overviewAnthropic Models API

Google

Gemini 3.1 Pro Preview

Frontier

Best current Artificial Analysis intelligence score and a top-three Arena text position, making it a strong general frontier benchmark.

AA intelligence

57

Arena overall

#3

Coding

#3

Creative

#2

Longer query

#3

Arena leaderboardArtificial AnalysisGemini model docs

Method stack

How this layer stays useful

Rule 1

OpenAI model availability is pulled from the live Models API when NUXT_OPENAI_API_KEY is present.

Rule 2

Anthropic model availability is pulled from the live Models API when NUXT_ANTHROPIC_API_KEY is present; otherwise the keyless OpenRouter public catalog keeps the lane current.

Rule 3

Provider lanes without a vendor API key fall back to the OpenRouter public model catalog (openrouter.ai/api/v1/models) before any hardcoded seed is used.

Rule 4

Arena and Artificial Analysis remain benchmark references for relative quality, while provider APIs decide whether a model is currently available.

Rule 5

Official vendor docs remain the source of truth for API availability, context windows and pricing.

Open Models

Open-weight leaders worth tracking

AA open-weight board + HF open LLM board

Z.ai

GLM-5 (Reasoning)

Current open-weight benchmark leader on Artificial Analysis, with a 200k context window and MIT licensing called out in the comparison pages.

AA open rank#1
AA intelligence50
Context200k
LicenseMIT
AA open weightsHF Open LLM board

Moonshot

Kimi K2.5 (Reasoning)

Still sits in the top open-weight tier for reasoning-heavy work and remains one of the clearest self-hostable alternatives to frontier closed models.

AA open rank#2
AA intelligence47
ProfileReasoning
TrackOpen weights
AA open weightsHF Open LLM board

Qwen

Qwen3.5 397B A17B

A top-three open-weight contender on Artificial Analysis that keeps showing up as a practical agentic and coding checkpoint to watch.

AA open rank#3
AA intelligence45
ProfileAgentic
TrackOpen weights
AA open weightsHF Open LLM board

Official Specs

Source-of-truth docs

OpenAI

GPT-5.6 Luna / GPT-5.6 Sol / GPT-5.6 Terra

Docs

Live OpenAI catalog lane for the newest GPT family available to the configured API key.

  • The OpenAI Models API returned GPT-5.6 Luna / GPT-5.6 Sol / GPT-5.6 Terra for the current key.
  • Official model docs remain the source of truth for context windows, pricing and intended use.
  • Keep benchmark placement separate from API availability: a model can appear in the catalog before third-party leaderboards fully settle.
OpenAI model docsOpenAI model API

Anthropic

Claude Opus 5 / Claude Sonnet 5

Docs

Official Claude lineup centered on Opus 4.7 for maximum agentic coding, Sonnet 4.6 for balanced production use and Haiku 4.5 for low-latency throughput.

  • The Anthropic Models API returned Claude Opus 5 / Claude Sonnet 5 for the current key.
  • Anthropic lists Claude Opus 4.7 as the most capable generally available Claude model and exposes it as claude-opus-4-7.
  • Published specs list a 1M context window and 128k max output for Claude Opus 4.7, with the same $5 / $25 per MTok headline pricing as Opus 4.6.
Claude models overviewAnthropic Models API

Google

Gemini 2.5 Pro / 2.5 Flash / 2.5 Flash-Lite

Docs

Official Gemini lineup optimized around deep reasoning, price-performance and ultra-fast throughput.

  • Google positions Gemini 2.5 Pro as the most advanced option for complex tasks.
  • Gemini 2.5 Flash is framed as the best price-performance choice for low-latency reasoning.
  • Gemini 2.5 Flash-Lite is the fastest and most budget-friendly multimodal option in the family.
Gemini model docs

Latest Launch

KickItLikeShika/qwen3-1.7b-dft-science

General • Created Jul 30

9 items in the latest lane

Hot Momentum

moonshotai/Kimi-K3

image-text-to-text • Trend 8.1K

9 items in the hot lane

Update Watch

0001AMA/ama-akuoko

General • Updated Jul 30

9 items in the updated lane

HF Frontier Feed

Fresh releases, crowd momentum and active model updates

Latest, hottest and recently updated Hugging Face models

Leaderboard Snapshot

Model Vault

추정 공식 카탈로그 메타데이터 기반 자체 추정치입니다. 실측 벤치마크가 아닙니다. Artificial Analysis · LMArena

1. GPT-5.6 Luna

ELO 2002

OpenAItext

Context

1M

Speed 추정

430

2. GPT-5.6 Sol

ELO 2002

OpenAItext

Context

1M

Speed 추정

430

3. GPT-5.6 Terra

ELO 2002

OpenAItext

Context

1M

Speed 추정

430

4. GPT-5.3 Codex

ELO 1961

OpenAIcode

Context

400K

Speed 추정

560

5. Claude Opus 5

ELO 1870

Anthropictext

Context

1M

Speed 추정

350

6. Claude Sonnet 5

ELO 1855

Anthropictext

Context

1M

Speed 추정

520

7. Gemini 3.6 Flash

ELO 1726

Googletext

Context

1M

Speed 추정

560

8. Gemini 3.5 Flash

ELO 1714

Googletext

Context

1M

Speed 추정

1.2K

Action playbook

Model evaluation steps

Validate moonshotai/Kimi-K3 against one benchmark workflow before adopting it.

Review the card and tags for KickItLikeShika/qwen3-1.7b-dft-science to see whether it is a genuine launch or a thin checkpoint wrapper.

Re-check 0001AMA/ama-akuoko when the lastModified lane moves, especially if it already fits your stack.

Use the hot, latest and updated lanes together before moving a model into production testing.

Task Mix

The most common task categories showing up across latest, hot and updated lanes.

General

HF lane density

17

text-generation

HF lane density

5

image-text-to-text

HF lane density

3

text-to-speech

HF lane density

1

object-detection

HF lane density

1