AI Model Launch Tracker

What we captured

Zhipu AI released GLM-5.3-Flash, an open-source MIT-licensed native multimodal model with a 1-million-token context window and 300 billion parameters. The company says it is served by more than 100,000 domestically produced chips.

• • Zhipu AI described GLM-5.3-Flash as the first native multimodal model in its GLM-5 series.
• • The model is released under the MIT license with a 1-million-token context window and 300 billion parameters.
• • Zhipu AI said its inference service used more than 100,000 domestically produced chips.
• • Published pricing is $0.15 per million input tokens and $0.50 per million output tokens.

Source evidence

Confidence 99% / Completeness 100%

What we captured

Z.ai released GLM-5.3-Flash with open weights, MIT licensing, a one-million-token context window, and API availability. The model is positioned as a lower-cost alternative to GLM-5.3 and was reportedly served on Chinese AI chips.

• • Z.ai described GLM-5.3-Flash as the first natively multimodal model in the GLM-5 series.
• • The model has 320 billion total parameters and 18 billion active parameters.
• • The model has a one-million-token context window and an MIT license.
• • Z.ai's API pricing is $0.15 per million input tokens and $0.50 per million output tokens.
• • Z.ai said the model ran on Chinese AI chips, with serving software built on SGLang.

Source evidence

What we captured

Leiolai launched leiolai-1 alongside a consumer app and developer API. The model runs inference across users' devices and offers an 11-million-token context window.

• • Leiolai launched leiolai-1 on 2026-08-27.
• • Inference runs across users' existing devices rather than data centers.
• • The model has an 11-million-token context window.
• • Fast and Research modes start at $0.01 per million input tokens and $0.02 per million output tokens.

Source evidence

Confidence 99% / Completeness 96%

What we captured

Alibaba’s Qwen released Qwen3.8-Flash with a 262,144-token default context window, expandable to 1 million tokens, lower stated training costs, and API pricing of 1 yuan per million input tokens and 3 yuan per million output tokens.

• • Qwen3.8-Flash is described as a multimodal model.
• • Its default context window is 262,144 tokens and can expand to 1 million tokens.
• • Alibaba said its training cost is about one-ninth that of Qwen3.7-Plus.
• • API pricing is 1 yuan per million input tokens and 3 yuan per million output tokens.
• • Open-source weights were also released for the related Qwen3.8-Flash-Next prototype.

Source evidence

What we captured

xAI made Grok 4.6 available to Azure customers through Microsoft Foundry in public preview, with a 500,000-token context window and published pricing of $2 per million input tokens and $6 per million output tokens.

• • Microsoft Foundry added Grok 4.6 on August 26, 2026.
• • The listing is a public preview.
• • The model has a 500,000-token context window.
• • Published pricing is $2 per million input tokens and $6 per million output tokens.

Source evidence

What we captured

Alibaba made Qwen3.8-Flash available through the managed QwenCloud API at published rates, with a default 1-million-token context and OpenAI- and Anthropic-compatible interfaces.

• • Alibaba Qwen announced the API launch on August 26, 2026.
• • QwenCloud lists rates of $0.16 per million input tokens and $0.47 per million output tokens.
• • The service provides a default 1-million-token context and supports OpenAI and Anthropic API specifications.

Source evidence

Confidence 99% / Completeness 96%

What we captured

Alibaba’s Qwen team released Qwen3.8-Flash-Next, an open-weight multimodal mixture-of-experts architecture preview for Qwen4, with production access offered as Qwen3.8-Flash through QwenCloud.

• • The model has 125B total parameters and activates 6B per token.
• • It includes a 51B-parameter N-gram embedding layer that can run in system RAM.
• • Native context is 262,144 tokens, expandable to one million with YaRN.
• • QwenCloud pricing is $0.16 per million input tokens and $0.47 per million output tokens.

Source evidence

What we captured

IBM released Granite 4.2 in 3B, 8B, and 30B sizes under Apache 2.0, with up to 512,000-token context and agentic capabilities in the 8B and 30B variants.

• • Released in 3B, 8B, and 30B sizes.
• • Supports up to 512,000-token context windows.
• • Includes thinking, non-thinking, and low-effort modes.
• • The 8B and 30B variants received agentic reinforcement-learning training for tool use, coding, and web search.
• • Supports OpenAI-format tool calling and runs on vLLM or SGLang.
• • Released under the Apache 2.0 license.

Source evidence

What we captured

Thomson Reuters launched Thomson, a Qwen-based legal language model trained with the company’s proprietary content and domain expertise. It is initially being deployed for document-review tasks in CoCounsel Legal, with a smaller open-weight, non-commercial version planned for Hugging Face.

• • Thomson Reuters says Thomson is built on Alibaba’s Qwen and trained with the company’s proprietary content and domain experts.
• • The company reports approximately $40 million in development spending over more than two years.
• • Thomson’s benchmark advantage over GPT-5.4 appeared only when both models could access Thomson Reuters content.
• • The initial deployment targets high-volume legal document-review subtasks in CoCounsel Legal.

Source evidence

What we captured

Thomson Reuters launched Thomson, including Thomson-1.0-Large, a proprietary legal model specialized from open-weight foundations with Thomson Reuters' legal and news content, expert data, and evaluations.

• • Thomson Reuters described Thomson as its first proprietary large language model.
• • The model was specialized from open-weight foundations using Thomson Reuters content, expert judgment, tools, and training methods.
• • Thomson Reuters reported approximately $40 million in development spending, excluding the undisclosed Safe Sign Technologies acquisition price.
• • Thomson-1.0-Large reportedly exceeded GPT-5.4 and Claude Sonnet 5 on an aggregate benchmark but trailed Claude Opus 4.8.

Source evidence

Confidence 99% / Completeness 79%