DeepSeek-V4.1-Flash
DeepSeek-V4.1-Flash is a 552B-parameter Mixture-of-Experts model with native visual understanding, which DeepSeek calls the smallest model in its new architecture family. It is served on the DeepSeek API under the model name deepseek-flash, with open weights under the MIT License.
Released
September 10, 2026
Type
multimodal
License
open-weight
MIT
Context
1,000,000 tokens
Pricing
Input
$0.30 / 1M tokens
Output
$1.20 / 1M tokens
List price on the developer's own pages, checked . Peak-hour price for cache-miss input. Off-peak is half ($0.15 input / $0.60 output). Cache-hit input $0.006 at peak, $0.003 off-peak. Peak hours are 01:00-04:00 and 06:00-10:00 UTC, Monday to Friday, excluding Chinese public holidays. The older IDs deepseek-v4-flash and deepseek-v4-flash-vision-exp are served by V4.1-Flash at this price.
Capabilities
Architecture
Parameters: 552B total (MoE); 8B active parameters for input (prefill), 16B for output (decode)
Sources
Every figure on this page was checked against these pages on . A figure we could not source is left out.
- DeepSeek API Docs: DeepSeek-V4.1-Flash release (2026/09/10) (opens in a new tab)
- DeepSeek API Docs: Change Log (opens in a new tab)
- DeepSeek API Docs: Models & Pricing (opens in a new tab)
- Hugging Face: deepseek-ai/DeepSeek-V4.1-Flash model card (opens in a new tab)
Spotted something out of date? Tell us.
Links
In the news: DeepSeek-V4.1-Flash and DeepSeek
MarkTechPost
Reflection AI Introduces Beam: A 501B Open-Weight MoE Model With 23B Active Parameters for Coding and Agentic Workloads
Reflection AI has introduced Beam, its inaugural open-weight model, on March 18, 2024. Beam is a sparse Mixture-of-Experts (MoE) architecture boasting a total of 501 billion parameters, with 23 billion active parameters per token. This model is specifically engineered for coding, reasoning, and agentic workloads. According to the Reflection AI team, Beam is positioned to directly compete with larger open models such as GLM 5.2, while demonstrating a significant advantage in inference compute efficiency, using 3 to 4 times less compute on reasoning benchmarks. Currently, Beam is not available for self-hosting but is undergoing final red-teaming. Early access is being managed through a waitlist on the Reflection platform. Beam functions as a general agent model, developed from the ground up by Reflection AI, with a primary focus on enterprise-level coding and agentic applications. Reflection AI frames Beam as a significant advancement in the Western open-weight model landscape. The research team acknowledges that while Kimi K3 may still lead in raw capability, Beam's key selling proposition is its efficiency during inference. Users can adjust a "reasoning effort" parameter, where lower settings yield concise answers and higher settings enable more extensive reasoning for complex tasks. This allows teams to tailor the model's effort to the specific difficulty of a task and their available compute budget. The pretraining of Beam involved an extensive dataset of 23.8 trillion tokens, sourced from the web, public repositories, and proprietary licensed datasets. The Reflection AI team reported that their curation process removed approximately 95% of raw internet tokens, while retaining about 1.8 trillion high-quality tokens that might have been discarded by conventional filtering methods. The model's architecture integrates local and global attention mechanisms with finely routed experts. Its load balancing strategy builds upon DeepSeek-V3's auxiliary-loss-free balancing, incorporating a cosine decay for expert-bias updates. This approach ensured that the most utilized expert experienced a load of only 1.04 times the average by the conclusion of pretraining. Throughout the 52 layers, residual norms were maintained within bounds through depth-based scaling, SandwichNorm, attention gating, and FP32 residual accumulation. The entire pretraining process was completed in under four weeks, utilizing 6,144 NVIDIA GB300 NVL72 GPUs. The reported goodput reached 92.3% near the end of the training period, with nine semi-automatic rewinds implemented. During mid-training, the effective context window was extended to 1 million tokens.
MarkTechPost
DeepSeek Harness v0.2 Brings Official Desktop Apps to Its Open-Source Agent Harness
DeepSeek has released an official desktop application for its open-source agent harness, DeepSeek Harness (dsh), with the v0.2 preview. Installers are available for macOS, specifically for Apple silicon, and for Windows 64-bit operating systems. This preview version is deployable immediately and can be downloaded from deepseek.com/harness or accessed via the command line using 'npx @deepseek-ai/dsh web'. DeepSeek has cautioned that compatibility-breaking changes are expected in future releases. The v0.2 preview enhances the harness's capabilities to target both everyday work and coding tasks. Key built-in features include preloaded, commonly used tools for office productivity and software development. A new plugin management interface allows users to install, configure, enable, and disable plugins through a dedicated page. For code and file review, a right-hand sidebar provides previews of generated files and code changes. The harness now supports document work, enabling users to input documents, spreadsheets, or PDFs and request the generation of charts or slides. An automation task plugin has been introduced, which schedules recurring prompts and includes a run history with editable frequency settings. The desktop build also integrates the dsh command-line interface, eliminating the need for separate Node.js or pnpm installations. Users can authenticate using a DeepSeek account or by providing an API key. According to the release candidate 1 (rc.1) notes, account-based models gain access to web search functionality without requiring an additional API key. The architecture of DeepSeek Harness is designed around a plugin-centric approach, where the model adapter, tool registry, and agent loop are all implemented as plugins, allowing for modularity and replacement of any layer. DeepSeek first open-sourced dsh in August 2026 under the MIT license. The harness runs on Cordis, as detailed in the research paper ‘A Programming Paradigm for Spatiotemporal Composability’. The harness is not restricted to DeepSeek's own models, with a provider guide available for integrating third-party providers and custom OpenAI-compatible endpoints. The project's GitHub repository has garnered significant attention, exceeding 240,000 stars and 24,000 forks. DeepSeek Harness also serves as DeepSeek's internal evaluation framework. The API changelog indicates that Code Agent benchmarks, such as the Terminal Bench 2.1, have been run using DeepSeek Harness in a minimal configuration. For instance, the DeepSeek-V4-Flash-0731 model achieved a score of 82.7 on Terminal Bench 2.1 within this setup. The harness supports four distinct session modes.
The Hacker News
Anthropic Says Seven China-Based AI Labs Ran Industrial-Scale Claude Distillation Attacks
Anthropic announced on Thursday that it has identified and disrupted industrial-scale illicit distillation attacks targeting its Claude AI model. These sophisticated attacks were traced back to seven distinct artificial intelligence laboratories based in China. The identified entities include prominent organizations such as Alibaba, Moonshot, DeepSeek, Z.ai (also known as Zhipu), and MiniMax. The attacks involved knowledge distillation, a legitimate machine learning technique where a larger, more capable AI model acts as a 'teacher' to train a smaller model. However, in this context, the labs were allegedly using Anthropic's powerful Claude models to illicitly train their own AI systems without authorization, bypassing Anthropic's terms of service and potentially infringing on intellectual property rights. Anthropic's security team detected these activities by monitoring for unusual patterns of access and data exfiltration that were indicative of large-scale, systematic attempts to replicate Claude's capabilities. The company stated that these distillation efforts were "industrial-scale," suggesting a significant investment of resources and computational power by the implicated labs. Such attacks aim to extract proprietary knowledge and model parameters from a teacher model to create a student model that mimics its performance, often at a fraction of the training cost and time. By engaging in this unauthorized distillation, the Chinese labs sought to gain a competitive advantage by rapidly developing advanced AI models based on Anthropic's research and development. In response to the discovery, Anthropic took immediate action to mitigate the threat. The company stated that it has taken steps to block the identified actors and prevent further unauthorized access and misuse of its Claude models. This disruption aims to protect Anthropic's intellectual property and maintain the integrity of its AI systems. The company emphasized that knowledge distillation is a valid technique when conducted ethically and with proper licensing, but the actions of these seven labs constituted a violation of their acceptable use policies. The incident highlights the ongoing challenges in securing advanced AI models against sophisticated intellectual property theft and unauthorized replication attempts, particularly in a competitive global AI landscape. Anthropic has not disclosed the specific versions of Claude that were targeted in these attacks, nor the exact methods used by the labs to extract the necessary data for distillation. However, the scale and systematic nature of the attacks suggest a coordinated effort. The company's proactive stance in identifying and reporting these activities underscores the increasing importance of AI security and the need for robust defenses against emerging threats. The involvement of multiple well-known Chinese AI companies also points to the intense competition and rapid development occurring within China's AI sector, where the pursuit of cutting-edge capabilities may sometimes lead to questionable practices. Anthropic's disclosure serves as a warning to other AI developers and a signal of the evolving security landscape in artificial intelligence.
MarkTechPost
DeepSeek AI Released DeepSeek-V4.1-Flash with 1M Context, FP4 KV Cache, and Cross-Layer Attention Reuse
DeepSeek AI released DeepSeek-V4.1-Flash, a multimodal Mixture-of-Experts large language model designed to address the input-heavy workload and significant KV cache strain associated with long-horizon agents and million-token contexts. The model boasts a 552 billion backbone parameter count and an additional 196 billion Engram parameters, enabling it to handle a 1 million token context window. During operation, DeepSeek-V4.1-Flash activates 8 billion parameters per token during the prefill phase and 16 billion parameters per token during decoding. A key innovation is its global KV cache footprint, which is reported to be 890 bytes per token. This represents a substantial reduction, being approximately one-quarter of the footprint of DeepSeek-V4-Flash and roughly 437 times smaller than that of DeepSeek-V1. The model's architecture incorporates a 40-layer backbone divided into a 20-layer causal encoder and a 20-layer decoder. Inspired by the YOCO architecture, the decoder does not compute its own global KV. Instead, per-layer projection weights derive the global KV from the final encoder hidden state. This design choice means that prompt tokens terminate at the encoder, which effectively halves the prefill compute required. The model also implements sliding-window attention (SWA) with a 128-token window in every layer. Decoder SWA states are reconstructed by replaying only the last 128 prompt tokens, a technique the research team refers to as Decoder SWA Bounded Replay. DeepSeek-V4.1-Flash utilizes a refined version of Compressed Sparse Attention (CSA2), building upon the CSA and Heavily Compressed Attention techniques used in DeepSeek-V4. This CSA2 implementation targets cache size reduction along the layer axis. Each CSA2 layer is assigned one of three static modes: 'Full', 'Reindex', or 'Reuse'. In 'Full' mode, a layer computes its own main KV and projects the indexer K from it, selecting the top 512 indices. The 'Reindex' mode reuses the main KV and indexer K from the last 'Full' layer but rescores them with its own indexer Q. The 'Reuse' mode goes further by reusing both the main KV and the latest Top-K indices, completely skipping the indexer computation. Each layer independently maintains its own main Q and SWA KV. The model features 18 CSA2 encoder layers. DeepSeek-V4.1-Flash is made accessible through open weights released under an MIT license. It is available via vLLM, SGLang, and Transformers integration paths on Hugging Face. Additionally, DeepSeek AI has announced a public API offering distinct reasoning tiers: low, high, and max. This release aims to provide a more efficient and deployable solution for handling extensive context windows in LLM applications.
The Hacker News
DeepSeek Harness Flaw Let AI Agents Disable Their Own File Sandbox Without Approval
A significant security vulnerability was discovered in DeepSeek Harness, an open-source tool developed by DeepSeek for executing AI coding agents on local developer machines. This flaw, identified and disclosed by security researchers, permitted a sandboxed AI agent to disable its own security sandbox with a single command, thereby circumventing intended safety protocols. DeepSeek Harness is designed to run an agent's commands within an operating-system-level sandbox. This isolation is crucial for preventing an AI agent, particularly one that might be processing untrusted files or code, from writing data or executing actions outside its designated workspace. The sandbox acts as a protective barrier, ensuring that the agent's operations remain contained and do not compromise the host system or other sensitive data. The vulnerability specifically allowed the AI agent to remove this critical security limitation by invoking a function within the tool's own web interface. This capability meant that an agent, once granted access to the harness, could effectively remove its own safety constraints without explicit user approval or oversight. The implications of such a flaw are substantial for the security of AI development workflows. If an AI agent can arbitrarily disable its sandbox, it could potentially access, modify, or exfiltrate sensitive files from the developer's machine, execute malicious code, or establish unauthorized network connections. This would undermine the fundamental purpose of sandboxing, which is to provide a secure environment for AI agents to operate, especially when dealing with potentially insecure code or data sources. The discovery highlights the ongoing challenges in securing AI systems, particularly those that interact directly with user environments and codebases. As AI agents become more integrated into development pipelines, ensuring the robustness of their containment mechanisms is paramount. The open-source nature of DeepSeek Harness means that the vulnerability could have affected numerous developers and projects that utilize the tool. Security researchers typically disclose such vulnerabilities after coordinating with the vendor to allow for a patch to be developed and deployed, thereby mitigating the risk to users. The specific details of the command or the web interface function that enabled this bypass were not fully elaborated in the initial reports, but the core issue revolved around the agent's ability to manipulate its own execution environment's security settings. This incident underscores the need for continuous security auditing and rigorous testing of AI development tools to prevent such critical breaches. The ability of an AI agent to self-modify its security posture represents a novel and concerning attack vector that developers and security professionals must now consider. The potential for an AI agent to gain unauthorized access or execute unintended actions due to a compromised sandbox is a direct threat to data integrity and system security in AI-powered development environments.