Seven AI models landed between August 10 and 15, 2026. Grok 4.6 went straight into Cursor and Grok Build. GLM-5.3 kept the GLM-5.2 base and concentrated its upgrade in post-training. Qwen3.8-27B paired dense multimodality with a high-teens-GB local quant. DeepSeek coupled API access to the exact downloadable revision; Qwen’s 2.4T checkpoint set the scale ceiling; Muse Glimmer and Nemotron arrived with local packages that include first-party runtime components.
Grok 4.6 went straight into coding tools
Grok 4.6 had the broadest launch surface of the group. xAI released it through its API, Grok Build, Cursor, OpenRouter, Vercel, and Cloudflare on August 12. GitHub Copilot followed two days later.
It accepts text and images, has a 500,000-token context window, and supports reasoning, function calls, and structured output. API pricing starts at $2 per million input tokens and $6 per million output tokens, then doubles once the input reaches 200,000 tokens.
GLM-5.3 puts the upgrade in post-training
GLM-5.3 reached all GLM Coding Plan users and ZCode on August 14. The standard API remains listed as coming soon, with downloadable weights promised about two weeks after launch.
Z.ai kept the GLM-5.2 base and attributes the improvements to post-training. The current model has 1 million-token context, a 128,000-token output limit, tool use, structured output, and low, high, and max reasoning effort.
Qwen3.8-27B has a 16.69 GiB local quant
Qwen3.8-27B is a dense 27.8B model with image and video input, Apache 2.0 weights, native 262,144-token context, an extension path to 1 million tokens, and optional thinking.
The full BF16 files occupy about 55.6 GB. Unsloth’s Dynamic GGUF reduces the UD-Q4_K_XL text artifact to 16.69 GiB, with vision handled by a separate projector. An RTX 5080 has 16 GB of VRAM and needs system-memory offload. An RTX 4090 or a Mac with 24 GB can hold the text quant with limited room left for context, vision, and the runtime. Unsloth published a guide for running locally.
DeepSeek V4 Pro 0813 is an API and an 893 GB checkpoint
DeepSeek-V4-Pro-0813 has two very different forms. The hosted API has 1 million-token context, up to 384,000 output tokens, thinking and non-thinking modes, tool calls, a Responses API, and an Anthropic-compatible endpoint. On August 16, its pricing moves to peak and off-peak rates, making request time and cache status part of the bill.
The MIT-licensed checkpoint is a 1.6 trillion-parameter mixture of experts with 49 billion active parameters. Its files total roughly 893 GB, and the published deployment example uses four GB300 systems. DeepSeek pairs a broadly accessible API with the exact downloadable revision at data-center scale.
Qwen3.8-2.4T-A95B is the giant
Qwen3.8-2.4T-A95B reaches 2.4 trillion total parameters with 95 billion active. It is text-only, keeps thinking enabled, starts with a native 262,144-token context window, and can extend to roughly 1 million tokens.
The checkpoint uses Qwen’s custom license, which adds conditions for some large commercial products and AI services. Hosted Qwen3.8 Max is a separate product with vision, optional thinking, built-in tools, and managed 1 million-token context.
Muse Glimmer ships a complete llama.cpp package
Meta Muse Glimmer is a 29.6B dense multimodal model under Apache 2.0. Meta publishes a 16.8 GB Q4 text model for 24 GB devices, a 1.4 GB vision projector, and an optional 1.6 GB DFlash drafter for faster generation.
The text model, image input, and speculative decoding come as separate first-party artifacts. The current package targets llama.cpp b10353 or later and lets applications load only the components they use.
Nemotron 3.5 Lightning activates only 3B parameters
NVIDIA Nemotron 3.5 Lightning has 30 billion total parameters and activates 3 billion per token. Its architecture mixes Mamba-2, mixture-of-experts, and attention layers.
NVIDIA pairs the model with NVFP4 and W4A16 checkpoints, hardware-specific vLLM recipes, and several speculative-decoding options. The current card also links official GGUF weights and an LM Studio route for local use. Its validated 1 million-token vLLM recipes remain tied to DGX Spark, H100 or H200, GB200, eight H100s, and Ampere; the RTX 5090 NVFP4 entry still has no published recipe.
Coding products and local runtimes moved into launches
Grok and GLM made coding products part of their releases. Grok arrived across Cursor, Grok Build, its API, and several gateways. GLM-5.3 started inside Coding Plan and ZCode while the standard API and weights waited. DeepSeek joined a live API to the exact 893 GB MIT-licensed checkpoint.
The local releases also came with different pieces of the runtime attached. Qwen3.8-27B gained high-teens-GB Unsloth quants, Muse Glimmer shipped its own llama.cpp projector and optional drafter, and NVIDIA attached hardware-specific recipes plus an official GGUF route to Nemotron. Nemotron activates 3 billion parameters per token while Qwen3.8-2.4T-A95B activates 95 billion, a 32-fold range behind the same 1 million-token target.