Qwen3.8-Flash
Qwen3.8-Flash is a multimodal Qwen model with a million-token context window, aimed at coding assistance, agentic workflows and visual understanding. It is the hosted version of Qwen3.8-Flash-Next, which Qwen describes as an experimental preview of the architecture that will underpin Qwen4.
Released
August 26, 2026
Type
multimodal
License
open-weight
Qwen Community License 1.0
Context
1,000,000 tokens
Pricing
Input
$0.15 / 1M tokens
Output
$0.47 / 1M tokens
List price on the developer's own pages, checked . QwenCloud list price (international). Implicit-cache input $0.016; explicit cache creation $0.2; explicit cache read $0.016. No input-length tiers shown.
Capabilities
Architecture
Parameters: 125B with 6B activated, plus 51B n-gram embedding and 4B MTP (open-weight base Qwen3.8-Flash-Next)
Sources
Every figure on this page was checked against these pages on . A figure we could not source is left out.
- QwenCloud: Model releases changelog (opens in a new tab)
- QwenCloud: Qwen3.8-Flash model page (opens in a new tab)
- Hugging Face: Qwen/Qwen3.8-Flash-Next model card (opens in a new tab)
- GitHub: QwenLM/Qwen3.8-Flash-Next README (news) (opens in a new tab)
- Alibaba: huggingface.co/Qwen/Qwen3.8-Flash-Next/blob/main/LICENSE (opens in a new tab)
Spotted something out of date? Tell us.
Links
In the news: Qwen3.8-Flash and Alibaba
MarkTechPost
Alibaba Qwen Introduces QwenGyre: Elastic Reinforcement Learning RL for XLong-Horizon Agents
Alibaba's Qwen team has introduced QwenGyre, an elastic reinforcement learning (RL) framework designed for training AI agents on extreme-long-horizon tasks. QwenGyre is an online reinforcement learning system that trains coding agents through unmodified, black-box harnesses, targeting rollouts that can run for hours and process close to 1 million tokens. The framework was tested on Alibaba's Qwen 3.6 122B and Qwen 3.8 2.4T models, with approximately 700,000 tokens per rollout. It operates on multi-node GPU clusters, with the 122B model utilizing 24 to 32 nodes and the 2.4T model using 192 nodes. The system employs two core ideas: it moves idle GPUs between rollout and training, and it transforms branching agent histories into bounded training samples. Performance tests showed QwenGyre to be 1.38x to 1.78x faster than Async and 1.21x to 1.85x faster than Colocate, achieving matched scores. The best performance was a 1.85x end-to-end speedup over Colocate on TerminalBench. QwenGyre addresses challenges in 'XLong-horizon' tasks, defined as agent runs spanning hours with hundreds of model and tool interactions, such as generating full repositories from specifications. Two primary problems at this scale are long rollout tails, where training waits for the slowest execution, and the complexity of branching agent histories, which inflate training costs by duplicating shared prefixes. The research team has not released public code for QwenGyre.
MarkTechPost
Microsoft AI Releases Microsoft-Decision-1: A Qwen3.5-9B Decision-Scoring Model
Microsoft has released Microsoft-Decision-1, a novel decision-scoring model designed for tasks such as routing, classification, verification, and agent control. Unlike generative AI models that produce text, Microsoft-Decision-1 is engineered to return a calibrated probability for each predefined answer option. This approach aims to provide outputs that software can act upon immediately, positioning decision models as a new category of AI. The model is post-trained from Alibaba's Qwen3.5-9B, though its exact parameter count has not been disclosed. Microsoft-Decision-1 features a substantial 32,768-token context window. It is currently accessible exclusively through hosted APIs, specifically within Microsoft Foundry and via OpenRouter on Azure. The model does not offer open weights or quantized variants, and details regarding its hardware specifications are not public. Microsoft has indicated plans to rebase future versions of its decision models on architectures from MAI and OpenAI. In performance evaluations conducted by Microsoft across 36 benchmarks, Microsoft-Decision-1 achieved the highest average accuracy, reaching 83.5%. This performance metric places it ahead of other models, such as Quyet-1.0-Large, which recorded an average accuracy of 81.9%. The model also demonstrated a p50 latency of 85 milliseconds. While its average accuracy was superior, its calibration score was 92.2%, ranking second to Quyet-1.0-Large's 93.1%. Microsoft-Decision-1 is strictly text-only and does not provide explanations for its outputs. The model's output format is JSON, supporting various question types including yes/no, multiple-choice, rating, classification, and the grading of AI responses and proposed agent actions. It also includes explicit abstention options, such as 'cannot tell'. The cost-effectiveness of Microsoft-Decision-1 is highlighted, with input tokens priced at $0.042 per million and free output tokens. However, a significant drawback noted is the closed nature of its weights. The training data for Microsoft-Decision-1 comprised public datasets acquired under Microsoft's Open Data process, supplemented with synthetic data. OpenRouter has stated that the model's weights are subject to continuous updates while maintaining a stable API structure. The decision model framework is intended to streamline AI integration into software workflows by providing direct, actionable scores rather than free-form text.
MarkTechPost
Alibaba Qwen Releases Qwen-Image-2.1-Turbo, an 8-Step 7B Image Model
Alibaba's Qwen team has introduced Qwen-Image-2.1-Turbo, a highly optimized checkpoint derived from their open-weight Qwen-Image-2.1 model. This new iteration dramatically enhances image generation and editing efficiency by reducing the denoising steps from the base model's 40 to a mere 8. This substantial acceleration, representing a five-fold improvement, is achieved while maintaining the same underlying 7 billion parameter visual generation architecture. For developers, this translates to significantly faster processing times and potentially lower computational costs when utilizing the model. The Qwen-Image-2.1-Turbo model builds directly upon the capabilities of Qwen-Image-2.1, offering identical performance in terms of generating high-resolution 2K images and supporting robust editing functionalities. The core of the Turbo model is a single-stream Diffusion Transformer (DiT) architecture, featuring 32 layers and the aforementioned 7B parameters. This architecture incorporates advanced block-causal attention mechanisms, which are crucial for its efficient operation. Complementing the visual generator is the Qwen3-VL 8B text encoder, responsible for processing textual instructions and any provided reference images. A notable technical advancement is the integration of a 64-channel RGBA Variational Autoencoder (VAE). This VAE, with its 16x spatial compression, enables native support for image transparency, a feature that expands the creative possibilities for users. The model's sampling process is managed by a Flow Matching scheduler, employing Euler discrete scheduling and dynamic shifting for optimal results. The sophisticated attention design facilitates prefix KV caching, a technique that allows for the intelligent reuse of text and reference-image context across multiple denoising steps, further contributing to the model's speed and efficiency. Qwen-Image-2.1-Turbo is designed to operate on CUDA GPUs utilizing BF16 precision through the widely adopted Diffusers library. While specific VRAM requirements for the Turbo version have not been officially published by the Qwen team, estimates for the base Qwen-Image-2.1 model suggest it can run on approximately 11 GB of VRAM when using GGUF quantization, or around 24 GB when employing INT8/FP8 precision. The performance benchmark for the base Qwen-Image-2.1 model is noteworthy; it achieved a score of 60.28 on the Qwen-Image-Bench, a metric that Alibaba's Qwen team reports as the leading score among currently available open-weight models. While Qwen-Image-2.1-Turbo is released as an open-weight checkpoint, it is important to note that it is distributed under a research license. This designation means that any commercial self-hosting or deployment of the model necessitates obtaining separate permission from Alibaba. To cater to developers who may not wish to self-host or require immediate access, Alibaba also provides a hosted API option for Qwen-Image-2.1-Turbo. During the image generation process, the model defaults to a Classifier-Free Guidance (CFG) value of 1, a common setting for balancing image quality and adherence to prompts.
MarkTechPost
Alibaba Qwen Releases Qwen-Audio-3.1-Realtime: A Full-Duplex Voice Model Trained to Think, Act, and Decide When to Speak
Alibaba's Qwen team has unveiled Qwen-Audio-3.1, an advanced audio processing stack comprising five distinct models designed for artificial intelligence agents. The centerpiece of this release is Qwen-Audio-3.1-Realtime, a full-duplex speech model engineered to facilitate seamless, real-time interaction for voice-enabled AI systems that require tool-calling capabilities. Concurrently, Qwen has implemented substantial price reductions across its audio offerings, with decreases of approximately 85% for the Realtime model, around 70% for its Text-to-Speech (TTS) capabilities, and up to 95% for its Automatic Speech Recognition (ASR) services. This new model is accessible as a managed API, with Qwen-Audio-3.1-Realtime-Plus now available on QwenCloud via WebSocket. No open-weight versions of the model have been announced. The QwenCloud deployment of Qwen-Audio-3.1-Realtime supports both text and audio as input and output modalities. It boasts a significant context window of 262,000 tokens, with a maximum input capacity of 245,000 tokens and a maximum output capacity of 16,000 tokens. The service operates under default limits of 60 requests and 100,000 tokens per minute. Pricing for the model on QwenCloud is set at $6.4 per 1 million audio input tokens and $0.8 per 1 million text input tokens. For output, both text and audio are priced at $24 per 1 million tokens, with text output not incurring additional charges beyond the token cost. Key functionalities integrated into the model include function calling, web search integration, generation of structured outputs, a context cache for efficient memory management, and fine-tuning capabilities. Complementing the Realtime model, Alibaba has also introduced Qwen-Audio-3.1-ASR-Flash-Filetrans, a model specifically designed for offline transcription of long audio files. This ASR model features advanced capabilities such as hot word detection, speaker diarization (separation), automatic punctuation insertion, and support for multilingual recognition, including various Chinese dialects. The pricing for Qwen-Audio-3.1-ASR-Flash-Filetrans is $0.15 per 1 million input tokens and $0.47 per 1 million output tokens. The underlying architecture of the Qwen-Audio-3.1 system involves two primary models that share a common Audio Encoder and Large Language Model (LLM) design. A full-duplex decision-making model is responsible for predicting the optimal action at any given moment: whether to continue listening, initiate speech, cease speaking, or resume listening. This is complemented by a speech-to-text model that transcribes the spoken content into text. Subsequently, a context-aware voice renderer converts this text into a streaming speech output, taking into account conversation history, vocal cues, and acoustic context to produce natural-sounding speech. The training methodology for this system is structured into three distinct layers: Think, Act, and Speak and Coordinate. The 'Think' layer utilizes M²-OPD Core-Cocktail SFT, a process that re-anchors the audio model to its source text LLM using extensive paired data, reportedly on the scale of millions of hours. This is followed by Multimodality OPD. In this stage, a Text Teacher model and a frozen Audio Reference model collaboratively score each token of the student model's generated trajectory. This approach is described as on-policy distillation, focusing on learning from the model's own actions rather than imitating pre-written responses. The training also incorporates domain experts to imbue the model with qualities like empathy and pragmatic understanding.
MarkTechPost
Alibaba Qwen Releases Qwen-Image-2.1: A 7B Open-Weight Model for Image Generation and Editing
Alibaba's Qwen team has released Qwen-Image-2.1, a unified open-weight model designed for both text-to-image generation and image editing tasks. This new model features a visual generation component with 7 billion parameters distributed across 32 single-stream Diffusion Transformer (DiT) layers. A single checkpoint now encompasses text-to-image generation, multi-reference editing, local editing capabilities, and the generation of transparent RGBA output. Qwen-Image-2.1 is available for research and evaluation purposes, with day-zero support for popular platforms including Diffusers, ComfyUI, vLLM-Omni, SGLang, and LightX2V. Commercial deployment of the model requires a separate license from Qwen. This release marks a significant reduction in size compared to its predecessor; the original Qwen-Image, launched in August 2025, was a 20 billion parameter model distributed under the Apache 2.0 license. The earlier iteration also required a separate Qwen-Image-Edit checkpoint for editing functionalities. Qwen-Image-2.1 consolidates both text-to-image generation and editing into a single model, approximately one-third the size of the original, which the Qwen team describes as the most balanced and cost-effective in the Qwen-Image series. It is important to note for capacity planning that the 7 billion parameter figure specifically refers to the diffusion transformer component; the overall pipeline also loads an 8 billion parameter Qwen3-VL encoder. The architecture of Qwen-Image-2.1, as detailed in its GitHub repository, comprises four key components. The Transformer consists of 32 layers with 7 billion parameters, employing a single-stream design with block-causal attention. A Qwen3-VL 8 billion parameter model serves as the text encoder, processing both text instructions and condition images into a unified representation. The model incorporates a 64-channel RGBA autoencoder with 16x spatial compression and native transparency support. For scheduling, it utilizes Flow Matching with Euler discrete scheduling and dynamic shifting. The model achieves its speed through a novel attention mechanism termed mixed-granularity attention. This approach employs a token-level causal mask for text tokens and a chunk-level bidirectional mask for image tokens within each image. Crucially, the condition prefix is positioned before the noisy latent and therefore does not attend to it. This means its keys and values remain fixed throughout the denoising steps. The model computes the text and input images only once, at the initial step, and reuses this prefix KV cache for all subsequent steps. This efficiency gain becomes more pronounced with an increasing number of reference images, explaining the model's claimed speed improvements for multi-image operations. Qwen-Image-2.1's capabilities include the native generation of RGBA images directly from text prompts, preserving transparency information.