Qwen3.8 Max Preview

The BenchLM.ai profile records that Qwen3.8 Max was released on August 3, 2026 as an open-weight reasoning model with a 1M-token context window, and assigns it a composite capability score of 78.7/100, ranking it 6th out of 230 tracked models. Its strongest category is Reasoning at the 100th percentile, supported by ca The same BenchLM record reports a measured throughput of 39 tok/s with a first-token latency of 53.79s, explicitly notes that no comparable first-party API token rate is published for this model, and dates its dataset to September 2, 2026. Because these speed and percentile numbers come from a third-party aggregator ra

A July 24, 2026 BigGo Finance report corroborates that Alibaba Cloud released the Qwen3.8-Max preview on July 19 as a native multimodal Mixture-of-Experts model with 2.4 trillion total parameters, described as Alibaba's first trillion-parameter-scale native multimodal model and accompanied by an explicit stated commitm Beyond the launch event itself, the piece places the preview in a broader corporate context, covering three Tongyi Lab restructurings in the 140 days since former technical lead Lin Junyang's departure, a pivot toward a commercially driven "Token Factory" model, and competitive pressure from KimiK3 and DeepSeek V4 as o

A third-party status page (qwen3coder.com), verified on July 21, 2026, documents the qwen3.8-max-preview hosted on Qwen Cloud under the Token Plan, with a 1M-token context window and Qwen-listed support for Thinking, function calling, and built-in tools; structured output is not listed in the current table. Qwen's laun The page explicitly distinguishes vendor-reported performance claims from independently verified facts, noting that the frontend improvement claim comes without a public benchmark table, a frozen model card, or a reproducible evaluation, so it is recorded as a Qwen statement rather than proven. It also confirms that Qw

The EvoLink blog post confirms that Qwen listed qwen3.8-max as its QwenCloud flagship on August 3, 2026, describing a native vision-language, 2.4-trillion-parameter Mixture-of-Experts model with hybrid thinking enabled by default, a 1M-token context window, Thinking, Function Calling, built-in tools, and Structured Out The same EvoLink article flags open-weight and license availability as not yet verified as released, recommends that integrators take the callable ID and live pricing from the product page rather than the docs slug, and advise requiring an account-level smoke test before production traffic. These caveats, combined with

On July 19, 2026, during the World AI Conference in Shanghai, Alibaba's Qwen team previewed Qwen3.8-Max-Preview, describing it as a 2.4 trillion-parameter multimodal model positioned as the next flagship in the Qwen family. The MarkTechPost write-up reports that Qwen's own benchmark table ranked the preview 'second onl The preview is a hosted model only: no downloadable checkpoint or reproducible evaluation accompanied the announcement, so the 2.4T parameter count and multimodal capability claim rest on Qwen's own materials pending first-party artifacts. For developers and tracking purposes, this piece functions as a launch record fo