Alibaba introduces Qwen3.8-Omni-Flash model with 1-million context window support

TaichuAI has made the ZDTaichu5.0-9B model publicly available. This multimodal foundation model has approximately 9 billion parameters and is designed for general visual understanding, spatial reasoning, agent tool usage, and embodied AI research.

The model's architecture combines the Qwen3.5-9B language backbone with the C-RADIOv4-H vision encoder. The model accepts text input, as well as one or more images and videos of any resolution, supporting a context length of up to 128 thousand tokens. The release, summarized by TMTPost on September 15th, focused on real-world spatial perception, transformations between different views, and task planning for embodied systems.

According to the public model card, TaichuAI demonstrates state-of-the-art results among approximately 10-billion general VLMs across a wide range of visual tasks, while also expanding its capabilities in spatial, embodied, and agentic scenarios. Noteworthy metrics in space and embodiment include ViewSpatial 62.50, MMSI-Bench 47.20, MindCube-tiny 78.27, ERQA 48.00, and RoboSpatial 56.00.

Regarding agent and instruction sets listed in the evaluation card notes, TAU2-Bench achieves a score of 87.70, and the average Claw-Eval score is 71.40, while IFEval shows 93.70. The entropy-prioritized adaptive recursive reasoning mechanism is described as a mechanism that allocates additional refinement steps for latent variables for more complex tokens.

The model's capabilities cover Optical Character Recognition (OCR) and document understanding, math on images, fine-grained 2D relationships, multi-view association, 3D scene and perspective perception, as well as tracking multiple images and videos within a long context window, including multi-step tool use planning. Actual tool execution remains the responsibility of the host application.

The spatial learning topics listed on the card include relative relationships, dense counting and framing, camera motion and depth ordering, egocentric versus allocentric views, and high-level capability definition and action planning for VLA-style adaptation.

The model weights are hosted on Hugging Face at TaichuAI/ZDTaichu5.0-9B under the NVIDIA Open Model License, while retaining the Apache-2.0 notices from Qwen3.5. A custom branch of vLLM 0.26.0 and a Docker image from TaichuAI are used for serving to provide OpenAI-compatible endpoints, with recommended sampling settings for spatial grounding and general tasks.

As with other vendor-released models, external laboratories should view benchmark scores as reported with the specified prompts and judges until independent reproduction results emerge. However, the combination of a mid-sized open multimodal checkpoint, a focus on space and agents, and ready vLLM packaging provides researchers with a concrete artifact for evaluation.

DeepSeek has announced the general availability (GA) release of the DeepSeek V4.1 Flash model. This model has been transitioned from a limited beta version to a production SKU based on the new Mixture-of-Experts (MoE) architecture with a Causal-Encoder-Decoder.

V4.1 Flash is positioned as the smallest member of this structural family and, according to the company, surpasses DeepSeek V4 Pro in performance metrics, cost, speed, and overall task completion time. This paves the way for the planned decommissioning of V4 Pro within the current week.

Architecturally, V4.1 Flash is an MoE model with 552 billion parameters and asymmetric activation: approximately 8 billion parameters are activated when reading input data, and about 16 billion are activated when generating output data. DeepSeek asserts that this separation reduces computational costs compared to denser models of similar overall size, which is particularly important for agent workloads that process large codebases, documents, and chat histories before providing concise decisions or tool calls.

The new pretraining combined with reinforcement learning post-training forms the basis of the GA checkpoint. Official data demonstrates high results on test sets such as GPQA Diamond, Codeforces-style programming, MathArena Apex, Terminal-Bench 2.1, and CyberGym. Furthermore, the stack provides KV cache compression compared to previous generations, reducing memory load during long sessions.

Image understanding functionality is now integrated directly into the main Flash API, rather than being an experimental add-on. Old V4 Flash and V4 Flash Vision Exp endpoints have been disabled; however, for compatibility, deepseek-v4-flash and deepseek-v4-flash-vision-exp are temporarily redirected to V4.1 Flash. Developers can pass public image URLs, encoded payloads, or API Files links for description, Optical Character Recognition (OCR), and graph reading. The context window remains around one million tokens with a larger maximum output budget, and it supports JSON output modes, tool calling, and thinking with a configurable effort level.

New commercial terms come into effect with this transition. New Flash tariffs (peak and off-peak) took effect at 12:00 Beijing Time on September 10, 2026, with the cost of input data hitting the cache during off-peak hours reduced by approximately 60% compared to previous Flash pricing—a direct advantage for agents repeatedly processing history. At 12:00 Beijing Time on September 14, DeepSeek will cease support for V4 Pro: requests to deepseek-v4-pro will be routed to V4.1 Flash and billed at Flash rates until the release of the future V4.1 Pro version. The GA model should be referred to as deepseek-flash.

In the consumer application, Quick, Expert, and Vision modes are merged into a single intelligent mode that determines complexity, activates vision upon receiving images, and scales the reasoning process without manual switching.

Thus, this is a formal reset of the architecture and product line, not just an expiring preview: it is a new MoE foundation, multimodal Flash as the primary work tool, cheaper traffic with high cache load, and a fixed calendar for Pro release and mode consolidation.