Ledger on GitHub. Hub here. Branches elsewhere.
August 2026: Three MoE releases, agent harnesses, and 750-token-per-second inference
August 2026: Three MoE releases, agent harnesses, and 750-token-per-second inference
Three mixture-of-expert architectures shipped from China inside two weeks, all with open weights and permissive licenses. OpenAI previewed throughput above 700 tokens per second on Cerebras silicon. Meta released a 30-billion-parameter Apache-licensed model built for agentic use. The batch arrived between early and mid-August 2026.
Alibaba Qwen 3.8
Alibaba announced Qwen3.8-Max on August 3, 2026. Reuters reported the architecture as a 2.4-trillion-parameter mixture-of-experts model with 95 billion active parameters and a 262,000-token context window.
The company also released Qwen3.8-27B, a 27-billion-parameter dense model, under an open license alongside the Max variant.
Both models appeared in the same launch window. No pricing or API details were included in the sources.
DeepSeek V4 Pro
DeepSeek released V4 Pro under MIT license on August 13, 2026. The model is a 1.6-trillion-parameter mixture-of-experts design with 49 billion active parameters and a 1-million-token context window.
Reuters reported API pricing at approximately $0.435 per million input tokens and $0.87 per million output tokens. The pricing was raised shortly after GA.
DeepSeek paired the model with DeepSeek Harness, an open-source agent framework for code generation and workflow automation. The harness ships as a separate repository and targets multi-step task execution.
Z.ai GLM-5.3 and Meta Muse
Z.ai shipped GLM-5.3 in mid-August. The model was noted for cybersecurity capabilities; one digest reported that it surfaced a vulnerability in Cursor, a code editor.
Meta released Muse Spark and Muse Glimmer during the same window. Muse Glimmer is a 30-billion-parameter model under Apache 2.0 license, designed for local deployment and agentic workflows.
No architecture details, context lengths, or pricing were provided in the sources for either GLM-5.3 or the Muse models.
OpenAI Ultrafast and Frontier
OpenAI previewed Ultrafast inference on Cerebras hardware. The company reported approximately 750 output tokens per second. The preview was published in an OpenAI index post with no GA date.
OpenAI also introduced Frontier, described as an enterprise AI-agent platform with shared context and system integration. No technical details, pricing, or availability timeline were disclosed.
Summary table
| Model | Parameters | Active | Context | License | Date | Price (per M tokens) | URL |
|---|---|---|---|---|---|---|---|
| Qwen3.8-Max | 2.4T MoE | 95B | 262K | Open | 2026-08-03 | — | Reuters |
| Qwen3.8-27B | 27B | — | — | Open | 2026-08 | — | Capital and Compute |
| DeepSeek V4 Pro | 1.6T MoE | 49B | 1M | MIT | 2026-08-13 | $0.435 / $0.87 | Reuters |
| DeepSeek Harness | — | — | — | Open | 2026-08 | — | AI Digest |
| GLM-5.3 | — | — | — | — | 2026-08 | — | AI Digest |
| Meta Muse Glimmer | 30B | — | — | Apache 2.0 | 2026-08 | — | AI Digest |
| OpenAI Ultrafast | — | — | — | — | 2026-08 | — | OpenAI |
| OpenAI Frontier | — | — | — | — | 2026-08 | — | Fruition |
Practitioner lines
Act: Download DeepSeek V4 Pro and Qwen3.8-27B weights. Test the Harness framework against your task library. Compare token latency on Ultrafast if preview access is available.
Watch: GLM-5.3 cybersecurity behavior and whether the Cursor vulnerability disclosure becomes public. Pricing changes on DeepSeek API after the August 13 adjustment. Meta Muse Glimmer agentic benchmarks when community evals land.
Ignore: Frontier feature lists until OpenAI publishes integration specs and a pricing sheet. Model-name hype without parameter counts or context-window measurements.