AI Model Launch Tracker

What we captured

Alibaba Cloud’s Qwen team released Qwen3.8-2.4T-A95B, a downloadable open-weight mixture-of-experts model with 2.4 trillion total parameters, 95 billion active parameters, adjustable reasoning effort, and self-hosting support.

• • The model has 2.4 trillion total parameters and 95 billion active parameters per token.
• • Its native context window is 262,144 tokens and can be extended to 1,010,000 tokens.
• • The model supports low, medium, and xhigh reasoning-effort settings.
• • The release supports self-hosted deployment using downloadable artifacts.

Source evidence

Confidence 99% / Completeness 91%

What we captured

Dealign.ai published Qwen3.8-27B-CRACK-GGUF, a 27B-parameter, safety-modified derivative of Qwen3.8-27B distributed as seven GGUF quantizations and a vision projector for local llama.cpp inference.

• • Repository named Qwen3.8-27B-CRACK-GGUF published and reproduced by dealign.ai
• • Described as a 27B-parameter derivative of Qwen3.8-27B with refusal behavior removed ("abliterated")
• • Repository metadata lists creation date as 2026-08-12
• • Distributed as seven GGUF quantizations (10.5 GB to 29.0 GB) and a 0.9 GB F16 vision projector
• • Recommended local build: 17.0 GB Q4_K_M quantization
• • Packaged for local multimodal inference with llama.cpp and provides OpenAI-compatible local server commands

Source evidence

Confidence 92% / Completeness 91%

What we captured

Nvidia announced Nemotron 3.5 Lightning (30B MoE, 3B active per token), released Aug. 11, with open weights under OpenMDW-1.1 and availability via Nvidia, Hugging Face, and ModelScope; paired with open-source NeMo Switchyard routing library.

• • Model: Nemotron 3.5 Lightning — 30B parameters, mixture-of-experts
• • Active parameters per token: 3B
• • Released Aug. 11 (article text)
• • Context window: up to 1,000,000 tokens
• • License: OpenMDW-1.1, open weights/training data/recipes
• • Availability: Nvidia, Hugging Face, ModelScope; NeMo Switchyard on GitHub

Source evidence

Confidence 95% / Completeness 95%

What we captured

Meta Platforms released Muse Glimmer, a 30-billion-parameter model available for developers to download and modify.

• • Meta Platforms announced the Muse Glimmer release on August 10, 2026.
• • The model has 30 billion parameters.
• • The weights are available for developer download and modification.

Source evidence

Confidence 99% / Completeness 87%

What we captured

Meta released the 30-billion-parameter multimodal Muse Glimmer model as open weights under Apache 2.0, with quantized configurations intended to run on 24 GB or 32 GB of memory and a 131,072-token context.

• • Meta announced the release on August 10, 2026.
• • The model has 30 billion parameters and is multimodal.
• • Meta says quantized configurations fit within 24 GB or 32 GB memory envelopes.
• • The release uses the Apache 2.0 license and supports a 131,072-token context.

Source evidence

What we captured

Abliteration.ai launched the abliterated-model-large-v2 API, a modified GLM-5.3 designed to reduce refusal behavior for cybersecurity and other sensitive use cases.

• • The model is derived from Z.AI’s GLM-5.3.
• • Modified weights are not publicly downloadable.
• • The service reduces refusal behavior and is marketed for offensive cybersecurity and red-team use.
• • Optional customer-configured policy controls are available.

Source evidence

Confidence 96% / Completeness 92%

What we captured

MiniMax released H3 on July 31, 2026, offering synchronized stereo-audio video generation through a hosted API and restricted downloadable open weights.

• • Released July 31, 2026.
• • Generates video with synchronized stereo audio in one pass.
• • Open-weight release published August 3, 2026, at 768p; 2K upscaling and Context-IR remain API-only.
• • Community license excludes the United States, European Union, United Kingdom, and South Korea from local deployment and use.

Source evidence

Confidence 98% / Completeness 91%

What we captured

Anthropic launched Claude Opus 5, positioning it as a near-frontier model with improved coding, knowledge-work, scientific-research, and agentic capabilities at lower cost than competing frontier models.

• • Anthropic announced Claude Opus 5 on July 24, 2026.
• • Anthropic says Opus 5 improves substantially over Opus 4.8 at the same cost.
• • Anthropic positions Opus 5 near Claude Fable 5's intelligence at half the price.
• • The model is available immediately and is the default on Claude Max.

Source evidence

Confidence 99% / Completeness 87%

What we captured

Moonshot introduced Kimi K3 on July 16, 2026 — an open-weights model described as 2.8 trillion parameters, multimodal, with up to a one-million-token context window; its license contains a commercial-use agreement requirement for high-revenue service providers.

• • Moonshot introduced Kimi K3 on July 16, 2026.
• • Kimi K3 described as 2.8 trillion parameters and native multimodal capabilities.
• • Kimi K3 supports a context window of up to 1,000,000 tokens.
• • Kimi K3 released with open weights published for public download.
• • Kimi K3 license clause: separate agreement required for model-as-a-service vendors with > $20M annual revenue.

Source evidence

What we captured

OpenAI launched the GPT-5.6 family (Sol, Terra, Luna) for general availability on July 9, 2026, introducing the ultra multi-agent capability and later reducing Luna and Terra prices on July 30, 2026.

• • "We’re launching the GPT‑5.6 family of models for general availability following our limited preview: our new flagship, Sol, alongside Terra, a balanced model for everyday work, and Luna, our most cost-efficient model."
• • Article date: "July 9, 2026 Product Release" announcing the GPT-5.6 launch.
• • "Update on July 30, 2026: OpenAI reduced the price of GPT‑5.6 Luna by 80% and GPT‑5.6 Terra by 20%."
• • "ultra is our highest-capability setting, coordinating multiple agents across parallel workstreams to finish complex tasks faster."
• • "GPT‑5.6 launches with our most robust safeguards to date, designed to be resilient against determined and adaptive misuse without broadly limiting legitimate work."
• • "GPT‑5.6 Sol sets a new high of 53.6 [on Agents’ Last Exam], eclipsing Claude Fable 5 (adaptive reasoning) by 13.1 points."

Source evidence