Jalapeño shows 1.9× efficiency; Qwen 3.8-Max hits 2.4 trillion parameters; DeepSeek V4-Pro reaches GA

OpenAI presented its first custom inference chip at Hot Chips in late August 2026. The chip, named Jalapeño, is an ASIC co-designed with Broadcom and optimized for large-scale inference workloads.

The published benchmarks show 1.5–1.9× work per watt compared to NVIDIA GB200 and GB300 platforms. Latency improvements range from 1.7–3.6× on general tasks and 2.1–4.1× on interactive workloads. The hardware is being deployed inside OpenAI’s own infrastructure rather than sold externally.

Small-volume rollout is scheduled for the end of 2026, with broader deployment planned for 2027. The design reflects a shift toward custom silicon for inference at scale, following the path already taken by Google, Meta, and Amazon in their respective infrastructures.

Alibaba’s Qwen 3.8-Max and Qwen 3.8-27B

Alibaba launched Qwen 3.8-Max in early August 2026 as its largest open-weight model to date. The model is a mixture-of-experts architecture with 2.4 trillion total parameters and approximately 95 billion active parameters during inference. Weights were announced for release within a week of the unveiling.

Alongside the flagship, Alibaba released Qwen 3.8-27B, a 27-billion-parameter dense vision-language model. The 27B model ships with a native context window of 262,000 tokens and supports up to 1 million tokens via YaRN extension. Both models are distributed under the Apache 2.0 license.

Qwen’s ecosystem crossed 3 billion downloads in 2026, with more than 2 billion downloads recorded on Hugging Face alone. The combination of scale, licensing, and multilingual capability has positioned Qwen as one of the most widely adopted open-weight model families globally.

DeepSeek V4-Pro and Harness

DeepSeek moved its V4-Pro model from preview to general availability on August 13, 2026. The model is accessible via app, web, and API. Pricing was published three days later on August 16: $1.32 per million input tokens and $3.96 per million output tokens.

The pricing structure places V4-Pro in the premium flagship tier, closer to GPT-4o and Claude Opus pricing than to open-weight or distilled alternatives. The move signals DeepSeek’s intention to compete on capability and reasoning performance rather than solely on cost.

Concurrent with the GA release, DeepSeek published DeepSeek Harness, an MIT-licensed agent framework built on the Cordis runtime. Harness is designed as a modular plugin system that turns large language models into autonomous agents capable of tool use, memory, and multi-step reasoning. The framework targets workflows similar to those enabled by Claude Code and OpenAI’s Assistants API.

DeepSeek Harness is positioned as open infrastructure rather than a proprietary agent layer. The decision to release under MIT reflects a strategy of ecosystem capture through open tooling while monetizing inference access.

Moonshot Kimi K3

Moonshot AI unveiled Kimi K3 in the same window, another large-scale open-weight model from China. K3 is reported at 2.8 trillion parameters, placing it in the same weight class as Qwen 3.8-Max. Details on active parameters, context length, and benchmark performance were not available in the briefing sources.

The timing of three major Chinese frontier releases—Qwen 3.8-Max, DeepSeek V4-Pro, and Kimi K3—within a two-week span in August 2026 indicates coordinated positioning ahead of the fall conference season and year-end procurement cycles.

Table: August 2026 model and hardware releases

Entity	Release	Type	Key numbers	Date	URL
OpenAI	Jalapeño	Inference ASIC	1.5–1.9× efficiency; 1.7–3.6× latency improvement	Aug. 2026 (Hot Chips)	https://openai.com/index/jalapeno-first-results/
OpenAI	Jalapeño deployment	Hardware rollout	End-2026 small volume; 2027 broader	Aug. 2026	https://techcrunch.com/2026/08/25/openais-jalapeno-chip-is-built-for-fast-inference-at-scale-benchmarks-show/
Alibaba	Qwen 3.8-Max	Open-weight MoE	2.4T parameters; 95B active	Early Aug. 2026	https://www.alibabacloud.com/en/press-room/alibaba-unveils-qwen3-8-max?_p_lc=1
Alibaba	Qwen 3.8-27B	Dense vision-language	27B; 262K native context; 1M via YaRN	Early Aug. 2026	https://www.we.inc/blog/qwen3-8-max-open-weights-build-apps
DeepSeek	V4-Pro GA	Flagship model API	$1.32/M input; $3.96/M output	Aug. 13, 2026	https://api-docs.deepseek.com/news/news260813/
DeepSeek	Harness	Agent framework	MIT license; Cordis runtime	Aug. 2026	https://pondero.ai/news/2026-08-14-deepseek-v4-pro-ga/
Qwen ecosystem	Adoption milestone	Distribution	3B+ downloads; 2B+ on Hugging Face	2026	https://www.bighatgroup.com/blog/china-ai-weekly-2026-08-22/
Moonshot AI	Kimi K3	Open-weight model	2.8T parameters	Aug. 2026	https://kafkai.ai/articles/ai-technology/ai-models-july-august-2026-overview/

Practitioner lines

Act: If you are building inference infrastructure at scale, model the Jalapeño efficiency numbers against your current GPU cost structure. The 1.5–1.9× work-per-watt improvement and 2–4× latency gains on interactive workloads represent a new baseline for custom silicon ROI. Wait for independent verification, but begin scenario planning now.

Watch: DeepSeek Harness is the first major open-source agent framework released by a Chinese frontier lab under a permissive license. If adoption mirrors Qwen’s download trajectory, Harness could become the de facto agent runtime for open-weight models within six months. Track plugin ecosystem growth and integration with non-DeepSeek models.

Ignore: The headline parameter counts (2.4T, 2.8T) without corresponding active-parameter, context-length, and benchmark data. MoE architectures make total parameter count a poor proxy for capability or cost. Focus on active parameters, memory bandwidth, and cost per token instead.