DeepSeek small agent model: 284B/13B MoE, 1M context, MIT weights, $0.14/$0.28 per 1M tokens.

WeChat AI's 617B MoE model with 23B active parameters, demonstrating the first sequence-length scaling at frontier scale via Hidden Decoding, powering WeChat.

Open-source 230B parameter MoE model optimized for multi-language coding, agentic workflows, and real-world development tasks with 74% SWE-bench performance.

NVIDIA's efficient open-weight 30B MoE hybrid model with 3B active parameters, 1M-token context, and single-GPU deployment for local reasoning and coding.

Previous WizardLM-2 8x22B snapshot.

Previous DeepSeek V3 snapshot. Not a GPT-4o / Claude 3.5 census.

DeepSeek V4 represents the next generation of DeepSeek's flagship AI models, building upon the success of V3 with enhanced capabilities in reasoning, multimodal processing, and agent-based interactions.

Open-source local inference engine for Qwen3.8-Flash-Next on Windows and Linux, with a chat interface and compatible agent APIs.

Command A+ is Cohere's first mixture-of-experts flagship: 218B total and 25B active parameters, 48 languages, vision input, Apache 2.0 open weights.

DeepSeek-V4.1-Flash is a MIT 552B multimodal MoE with 8B/16B activation, 1M context, and about 1/4 the KV cache of V4-Flash.

DeepSeek-V4-Flash-Vision-Exp is the first V4 multimodal experiment: MIT vision weights on Flash, 305B params, screenshots plus text agents.

Fireworks Research rebuilt Kimi K3 to think less: same benchmark quality with roughly 40 percent fewer tokens, a 1M context, and $3 input per million tokens.

Google DeepMind's open-weight 25.2B MoE model with 3.8B active parameters, 256K context, multimodal input, and a commercially permissive Apache 2.0 license.

Zhipu's first natively multimodal GLM (confirmed as the 'Ox Alpha' stealth model): 320B MoE with 18B active, 1M context, MIT open weights at $0.15/M input.

Tencent's next-generation open-weight MoE flagship: 770B total params, 49B activated per token, 1M-token context. Apache 2.0, Gated DSA attention, and blind-eval scores edging out GLM 5.3 and Kimi K3.

Kwaipilot's open-weight agentic coding model: 35B MoE, 3B active per token, Qwen3.6 base, Apache 2.0, 262K context, top PinchBench tool-use score.

Aleph Alpha’s open-weight reasoning model for German and English, with 78B total parameters, tool calling and long-context inference.

Poolside's open-weight 118B MoE coding model with 8B active parameters, a 1M-token context window, native interleaved reasoning, and an OpenMDW-1.1 license.

Ling-3.0-flash-VL is inclusionAI's MIT multimodal MoE: 124B total, 5.5B active, with image and video.

Meituan's open-weight MoE model: 1.6T total / ~48B active params, 1M context, MIT. Strong on coding and agentic tasks (2026-08-25).

Xiaomi's MiMo-V2.6 open-weights series pairs a 1.02T-parameter omni MoE with a 309B Flash tier, both MIT licensed with 1M-token context.

mini-AGI is a byte-level model that trains from scratch on an 8GB GPU and keeps learning from one stream without forgetting what it knew.

Alibaba's next-generation Qwen 4 flagship is in training, with a roadmap that projects the Qwen 4.5 and Qwen 5 series to scale to 5-10 trillion parameters.

Alibaba's open-weight architecture preview of Qwen4: 125B multimodal MoE with 6B active plus a 51B N-gram table, 262K native context, at $0.16/M input.