🏠 Home
FRONTIER MODELS · RELEASE TRACKER

🚀 LLM Model Release Tracker

Every frontier-scale model release worth tracking — specs, timing, and what it means.

updated 2026-08-13 · Notion · Frontier Model Releases (be1366fa-33a8-4eda-8b83-6abf18ce78d1)

24
models tracked
14
Chinese labs (14/24)
14/23
released @ 1M context
4
releases in Aug '26

💡 Key insights

1. 9 of 16 released models ship 1M context — only older Claude Opus 4.7/4.8 and MiniMax M2.7 lag. Context length stopped being a differentiator.
2. MoE everywhere: DeepSeek V4-Pro 1.6T total / 49B active (3%), MiniMax M3 428B/23B, Kimi K3 2.8T/16B (0.6%). Inference cost ≈ active params — that asymmetry is the whole MoE business model.
3. 16 releases in ~5 months; 4 landed in the first 2 weeks of August alone (Qwen 3.8 Max, Grok 4.6, DeepSeek V4-Pro, Gemini 3.7 Flash).
4. Claude and Gemini publish no param counts ('not public') — they compete on capability, not spec sheets. Open labs publish everything.
5. 10 of 17 rows are Chinese labs (DeepSeek ×3, Kimi ×3, Qwen, GLM, MiniMax ×2) — and GLM-5.3 is pending.
6. No Mistral, no Llama/Meta, no xAI mid-tier, no Apple. The table tracks who's actually shipping frontier-scale.

🗓️ Release timeline

Jan Feb Mar Apr May Jun Jul Aug Kimi Kimi K2.5 Kimi K3 Kimi K2.7 Code Kimi K2.6 Claude Claude Opus 4.6 Claude Opus 5 Claude Opus 4.8 Claude Opus 4.7 MiniMax MiniMax M2.5 MiniMax M3 MiniMax M2.7 Qwen Qwen 3.5 (397B) Qwen 3.8 Max Gemini Gemini 3.1 Pro Gemini 3.7 Flash GPT GPT-5.4 GPT-5.3 Codex GPT-5.6 GLM GLM-5.3 (expected) GLM-5.2 Grok Grok 4.6 DeepSeek DeepSeek V4-Pro (0813) DeepSeek V4-Flash DeepSeek V4-Pro

📊 MoE efficiency: total vs active params

Qwen 3.5 (397B) 397B / 17B active MiniMax M2.5 229B / 10B active Kimi K2.5 1.0T / 32B active Qwen 3.8 Max 2.4T Grok 4.6 1.5T Kimi K3 2.8T Kimi K2.7 Code 1.0T Kimi K2.6 1.0T DeepSeek V4-Pro (0813) 1.6T / 49B active DeepSeek V4-Flash 284B / 13B active DeepSeek V4-Pro 1.6T / 49B active MiniMax M3 428B / 23B active MiniMax M2.7 229B / 10B active GLM-5.2 753B ■ total params (dim) · ■ active params (accent) — log scale

📋 All models

Model Family Release Date Parameters Context Status
GPT-5.4GPT2026-03-05not public1MReleased
Gemini 3.1 ProGemini2026-02-19not public1MReleased
Qwen 3.5 (397B)Qwen2026-02-16397B17B active262KReleased
MiniMax M2.5MiniMax2026-02-12229B10B active192KReleased
GPT-5.3 CodexGPT2026-02-05not public400KReleased
Claude Opus 4.6Claude2026-02-05not public1MReleased
Kimi K2.5Kimi2026-01-271.0T32B active256KReleased
Qwen 3.8 MaxQwen2026-08-032.4T1MReleased
Grok 4.6Grok2026-08-071.5T (V9 base)1MReleased
Gemini 3.7 FlashGemini2026-08-13not public1MReleased
GPT-5.6GPT2026-07-09not public1MReleased
Kimi K3Kimi2026-07-162.8T total / 16 of 896 active1MReleased
Kimi K2.7 CodeKimi2026-06-12~1T (MoE)256KReleased
Kimi K2.6Kimi2026-04-21~1T (MoE)256KReleased
DeepSeek V4-Pro (0813)DeepSeek2026-08-071.6T49B active1MReleased
DeepSeek V4-FlashDeepSeek2026-04-24284B13B active1MReleased
DeepSeek V4-ProDeepSeek2026-04-241.6T49B active1MReleased
Claude Opus 5Claude2026-07-24not public1MReleased
Claude Opus 4.8Claude2026-05-28not public200KReleased
Claude Opus 4.7Claude2026-04-16not public200KReleased
MiniMax M3MiniMax2026-06-01428B23B active1MReleased
MiniMax M2.7MiniMax2026-03-18229B10B active200KReleased
GLM-5.3GLM—TBDTBDExpected
GLM-5.2GLM2026-06-13753B (MoE)1MReleased
Data: Notion · Frontier Model Releases (be1366fa-33a8-4eda-8b83-6abf18ce78d1) · verified against lab announcements & Artificial Analysis. Page regenerated biweekly from the Notion table — Notion remains canonical. Built by Heidi.