Last updated on Jun 26, 2026, 3:40 PM
| 🤖 Model | 🏢 Developer | Video | 📅 Release Date | 📜 License | 🧠 Architecture | ⚡ Active Params | 📚 Context Window | 👁️ Modalities | 🎯 Use Case Fit | 💻 SWE-Bench | 🏆 Top Benchmark | ☁️ Deployment | 💲 Run Cost Tier | ⭐ Key Strengths | 🔗 HuggingFace / Link |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Gemma 4 | Google DeepMind | https://www.youtube.com/watch?v=7LEvSOiTWZk | 4/1/2026 [no-highlight] | Apache 2.0 | Dense + MoE | 31B / ~4.5B (edge) | 256K | Text, Image, Audio | Coding, Agents, Reasoning, Multilingual | 65% | AIME, GPQA, LiveCodeBench | API + Self-host | Low | Reasoning, coding, agentic tasks, efficient hybrid attention | https://huggingface.co/google/gemma-4 |
| Qwen3.5-397B-A17B | Alibaba Qwen | https://www.youtube.com/watch?v=hAFbrO-VNVw | Feb 2026 | Apache 2.0 | Hybrid MoE + Linear Attention | ~17B active | 262K (up to 1M+) | Text, Image, Video, Documents | Coding, Agents, Long-doc, Multilingual | 68% | Instruction following, reasoning, coding | API + Self-host | Low | Native vision-language, agentic workflows, math/coding, 201 languages | https://huggingface.co/Qwen/Qwen3.5-397B-A17B |
| MiniMax-M3 | MiniMax | https://youtu.be/yWXK6zu_kGE?si=8AQ6uf2pvNe_i-Cy | Jun 2026 | MiniMax Community | MoE | 23B active | 1M | Text, Image, Video, Computer Use | Coding, Agents, Vision | 74% | SWE-Bench style, long autonomous runs | API | Mid | Frontier-level coding, native multimodality, long-horizon agents | Fireworks / OpenRouter |
| Kimi-K2.6 | Moonshot AI | https://www.youtube.com/watch?v=LSfpwaujqLQ | 2026 | Modified MIT | MoE | ~32B active | 256K | Text, Image, Video | Coding, Agents, Long-doc | 73% | Long-horizon coding, agent swarm performance | API + Self-host | Mid | Long-horizon coding, agent swarms (300+ sub-agents), proactive autonomous agents | https://huggingface.co/moonshotai |
| DeepSeek-V4 Pro | DeepSeek | https://www.youtube.com/watch?v=UcSPD64453U | 2026 | MIT | MoE | 49B active | 1M | Text | Coding, Reasoning, Long-doc | 72% | SimpleQA-Verified, LiveCodeBench | API + Self-host | Mid | Adaptive think modes, world knowledge, efficient long-context reasoning | https://huggingface.co/deepseek-ai |
| DeepSeek-V4 Flash | DeepSeek | https://www.youtube.com/watch?v=2FwWRVHhdiE | 2026 | MIT | MoE | 13B active | 1M | Text | Coding, Reasoning | 65% | LiveCodeBench, fast inference | API + Self-host | Low | Fast, efficient variant of V4; strong coding at low cost | https://huggingface.co/deepseek-ai |
| GLM-5.2 | Zhipu AI (Z.ai) | https://www.youtube.com/watch?v=_qGxgZmSE4Y | Jun 2026 | MIT | MoE + IndexShare | ~40B active | 1M | Text | Coding, Agents, Long-doc | 76% | SWE-Bench Pro, Terminal-Bench | API + Self-host | Mid | Long-horizon agentic reasoning, repository-scale work, multiple thinking effort levels | https://huggingface.co/zai-org/GLM-5.2 |
| MiMo-V2.5-Pro | Xiaomi MiMo | https://www.youtube.com/watch?v=popj1l3LA-I | Apr–May 2026 | MIT-like (permissive) | MoE + Hybrid Attention + MTP | 42B active | 1M | Text, Image, Video, Audio | Coding, Agents, Vision, General | 71% | Agentic coding, complex SE trajectories | Self-host | Low (self-host) | Agentic capabilities, complex software engineering, instruction following at 1M context | https://huggingface.co/XiaomiMiMo/MiMo-V2.5-Pro |
| Qwen3-Coder-Next | Alibaba Qwen | https://www.youtube.com/watch?v=Rrqlt8Mm1oc | Feb 2026 | Apache 2.0 | Hybrid MoE | ~3B active | Long coding context | Text | Coding | 70.60% | SWE-Bench Verified ~70.6% | API + Self-host | Low | Specialized for coding agents, tool use, local development workflows | https://huggingface.co/Qwen/Qwen3-Coder-Next |
| Gemma 4 E4B (Edge) | Google DeepMind | https://www.youtube.com/watch?v=j1EchjiViOs | Apr 2026 | Apache 2.0 | Small MoE / Edge optimized | ~4.5B | 256K | Text, Image, Audio | Coding, Reasoning, On-device | 48% | Competitive for size on reasoning/coding | Self-host / On-device | Very Low | On-device multimodal reasoning and coding; efficient edge deployment | https://huggingface.co/google/gemma-4 |
| Un-0 | Unconventional AI | — | Jun 2026 | Open (verify) | Physics-based (coupled oscillators) | Not disclosed | N/A | Image generation | Vision, Research | N/A | Novel non-neural ImageNet 64x64 | Self-host / Research | Low (experimental) | Physics-inspired image generation, power-efficient via physical dynamics | Unconventional AI announcement |
| Wan 2.2 / Wan-Video | Wan-Video Team | https://www.youtube.com/watch?v=XGB4qBkCFSM | 2026 | Apache 2.0 | MoE Diffusion | Not specified | N/A (video length) | Text-to-Video, Image-to-Video, Editing | Video generation | N/A | Leads Wan-Bench; cinematic quality | Self-host | Mid (GPU required) | Cinematic video, long prompts (5000+ chars), 9-grid image input, first/last frame control | https://github.com/Wan-Video |
| LTX-2.3 | Lightricks | https://www.youtube.com/watch?v=W-PIgkRWJOc | Mar 2026 | Open weights (permissive) | Diffusion + Audio | 22B | N/A (video focused) | Text-to-Video, Audio | Video generation, Audio | N/A | Strong open-weights video; 4K at 50fps | Self-host | Mid (GPU required) | High-resolution video (4K), native audio generation, portrait/vertical video, fast inference | https://github.com/Lightricks/LTX-2 |