Qwen3.8-27B
Qwen3.8-27B is an open-weight 27-billion-parameter dense native vision-language model that understands images and videos, with thinking on by default and adjustable reasoning effort. It is released under Apache 2.0.
Released
August 19, 2026
Type
multimodal
License
open-weight
Apache 2.0
Context
262,144 tokens
Pricing
Input
$0.50 / 1M tokens
Output
$3.00 / 1M tokens
List price on the developer's own pages, checked . QwenCloud list price (international). Implicit-cache input $0.1; explicit cache creation $0.625; explicit cache read $0.05.
Capabilities
Architecture
Parameters: 27B (dense)
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)
- Hugging Face: Qwen/Qwen3.8-27B model card (opens in a new tab)
- QwenCloud: qwen3.8-27b model page (opens in a new tab)
Spotted something out of date? Tell us.
Links
In the news: Qwen3.8-27B 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
Meet the Underdog Saluki 27B: A 2-bit Qwen3.8-27B That Beats the Original at Tool Calling
Underdog, a division of Conway Research, has released Saluki 27B, a highly compressed 2-bit GGUF model derived from Qwen3.8-27B, under the permissive Apache 2.0 license. This new model is specifically engineered for on-device AI assistants and achieves a remarkable file size of 7.89 GB, a substantial reduction from the 54 GB required by the full BF16 version of Qwen3.8-27B. The primary objective of this compression was to preserve and enhance tool calling capabilities, a critical function that transforms a standard chat model into a more versatile AI agent. Developers can now leverage a 27B-parameter agent model that operates seamlessly within the stock llama.cpp framework, enabling broader accessibility and deployment on consumer hardware. Saluki 27B is built upon a multi-layered approach to quantization and fine-tuning. The foundation is Qwen3.8-27B, a dense 27 billion parameter model developed by the Qwen team, featuring 64 layers and incorporating Gated DeltaNet linear attention alongside gated attention, with native support for up to 262,144 tokens. The second layer of development involved integrating ISTA-DASLab’s Qwen3.8-27B-GSQ-RCO-GGUF, which employs Group-wise Scalar Quantization (GSQ) to learn precise low-bit scalar grids for each tensor and Residual-based Channel Optimization (RCO) to assign quantization types within a fixed size budget. ISTA's smallest variant, IQ2_XS, typically occupies 8.4 GB at 2.50 bits per weight. Underdog's proprietary third layer further refined this process, shrinking the file to 7.89 GB and specifically optimizing for tool calling performance. This final pass resulted in a file named IQ2-mix, which includes an imatrix tag, though Underdog has not disclosed the complete methodology behind this optimization. Performance evaluations indicate that Saluki 27B maintains a strong overall capability, achieving 96% of the performance of the full Qwen3.8-27B model across nine benchmarks. Its most significant advantage lies in parallel tool calls, where it surpasses the original model by a considerable margin, handling 42 concurrent calls compared to the full model's 35, representing a 120% retention rate in this specific function. However, the aggressive quantization does lead to a noticeable decrease in performance on certain complex reasoning tasks. For instance, in the AIME 2025 benchmark, which focuses on competition mathematics, Saluki 27B achieved a score of 79.2%, compared to 96.7% for the full model, indicating approximately an 82% retention rate. This suggests a trade-off where enhanced agentic capabilities are prioritized over performance in advanced mathematical and multi-step reasoning problems, with a drop of 12 to 18 percentage points in these areas. For developers and users seeking efficient, on-device AI agents, Saluki 27B presents a compelling option. Its ability to run on stock llama.cpp, with full GPU offload capabilities, makes it highly accessible. Furthermore, an optional vision add-on, available in 629 MB or 928 MB sizes, further expands its utility by enabling visual understanding. The model's success in outperforming the larger, uncompressed original model in critical agent functions like tool calling, while fitting into a fraction of the storage space, highlights the advancements in efficient AI model deployment. The trade-off in specific reasoning benchmarks is a key consideration for applications requiring high precision in areas like competitive mathematics or complex problem-solving.
MarkTechPost
Cloudflare Releases Clef and Clef-flash: Open-Weight Decision Models That Return Typed Probabilities Instead of Text
Cloudflare has released Clef and Clef-flash, two new open-weight decision models developed by its Workers AI team. Unlike traditional large language models (LLMs) that generate text token by token and require parsing, these models are designed to process an input state and a schema of typed questions, returning precise probabilities for each allowed answer without generating free-form text. This approach aims to provide greater predictability and control in AI outputs. Both Clef and Clef-flash are released under the Apache 2.0 license, making them open-weight and available for self-hosting. They are also compatible with TypeSafe AI’s Jev API, a system designed for structured AI outputs. These models are immediately deployable on Cloudflare’s Workers AI platform, and their weights are accessible on Hugging Face. A decision model's core function is to answer a predefined set of questions about an input, rather than generating novel text. Clef supports three primary question types: 'noul' for yes/no questions, returning the probability of a 'yes' answer; 'choice' for selecting one option from a list, providing per-option probabilities and a confidence score; and 'score' for rating against an ordered rubric, yielding a probability-weighted score. On the Workers AI platform, a single request can accommodate up to 64 questions and as many as four images. Clef is a post-trained version of Qwen3.8-27B, while Clef-flash is based on Qwen3.5-9B. Both models retain the vision encoder from their respective Qwen backbones. The inference process involves two stages. Initially, the backbone performs a single prefill-only pass over the input state and questions. Subsequently, a smaller transformer, termed the joint schema head, processes the final hidden states. This head is responsible for routing relevant evidence to each question, enabling cross-attention between fields, and jointly scoring all potential options. A per-question softmax function then converts the raw logits into probabilities. During training, the backbones were frozen, and the routing head was jointly optimized using rank-256 low-rank adapters. The training loss incorporates label-smoothed cross-entropy alongside a Brier loss for improved calibration. An additional objective, Reinforcement Learning for Calibrated Decisions (RLCD), is employed to provide partial credit for adjacent ordinal choices, further refining the model's ability to make calibrated decisions. The development of Clef and Clef-flash signifies a move towards more structured and interpretable AI outputs, addressing a key challenge in current LLM deployments where output parsing and validation can be complex. By returning typed probabilities, these models offer a more direct and reliable way to integrate AI decision-making into applications that require deterministic outcomes or precise confidence assessments. This contrasts with the generative nature of traditional LLMs, which can sometimes produce outputs that are ambiguous or require significant post-processing to extract actionable information. The open-weight nature of Clef and Clef-flash also encourages community development and adoption, potentially leading to new use cases and further advancements in decision-oriented AI.