DeepSeek-V4.1-Flash incorporates Compressed Sparse Attention 2 and an FP4 global key-value cache that reduces memory footprint to 890 bytes per token

Overview Debates Related claims Topics Sources

Chinese AI company MiniMax released MiniMax M3.1 Flash on October 19, 2026, featuring 428 billion total parameters, about 23 billion active parameters in a mixture-of-experts architecture, and a 1,000,000-token context window.

MiniMax reported that MiniMax M3.1 Flash operates about three times faster than MiniMax M3, with median output speeds consistently exceeding 170 tokens per second.

MiniMax M3.1 Flash API pricing is set at $0.15 per million input tokens and $0.60 per million output tokens, with cached input reads priced at $0.015 per million tokens.

MiniMax confirmed that an earlier build of MiniMax M3.1 Flash operated anonymously under the name Space Bunny, processing over 200 trillion tokens prior to its official launch.

MiniMax reported that MiniMax M3.1 Flash can reach peak generation speeds exceeding 200 tokens per second.

LiteLLM, developed by BerriAI, provides a shared interface for LLM providers including OpenAI, Anthropic, and Google Gemini, receiving approximately 23 million weekly downloads.

The complaints allege Meta Platforms Inc. operated a book-scanning program called Project Vellum starting around mid-2025 that purchased and scanned physical books to train its Muse Spark and Muse Glimmer AI models.

The authors' complaints allege Meta Platforms Inc. downloaded e-books from pirate shadow libraries, including LibGen, Z-Library, Anna's Archive, and the Books3 dataset, to train its large language models.

Artex was developed by Chinese engineer Li Puhua in July 2026 as a penetration testing tool that automates security assessments using large language models.

Wiz researchers noted that while AI helped identify the ancient PHP vulnerabilities CVE-2026-109618 and CVE-2026-109619, large language models also lower the barrier for attackers to adapt exploits across different server configurations

Chinese technology companies seek to use Nvidia RTX PRO 5500 workstation cards primarily for artificial intelligence inference tasks rather than model training

Chinese AI startup DeepSeek trained the debut version of its artificial intelligence model for approximately $6 million, compared to roughly $100 million spent by OpenAI to train GPT-4

A hacking campaign by a Chinese-speaking hacker deployed autonomous AI agents using a mixture of DeepSeek and Anthropic's Claude Code models

On October 17, 2026, DeepSeek released DeepSeek-V4.1-Flash-1017, an updated checkpoint of its V4.1-Flash artificial intelligence model accessible via DeepSeek's website, mobile app, and API

DeepSeek-V4.1-Flash-1017 completes tasks faster and produces shorter outputs than the earlier September 2026 release of DeepSeek-V4.1-Flash

DeepSeek-V4.1-Flash architecture utilizes a 552-billion parameter backbone and 196-billion Engram parameters, activating 8 billion parameters per token during prefill and 16 billion during decoding

On the Terminal-Bench 2.1 benchmark, Reflection AI's Beam model achieved a score of approximately 80% while generating roughly 25,000 tokens per attempt, compared to Z.ai's GLM-5.2 generating roughly 80,000 tokens

Reflection AI provided pre-configured Beam integrations for popular serving engines and custom fused kernels to speed up processing

The 8-bit and 4-bit quantized versions of Reflection AI's Beam model can run on an eight-GPU server powered by Nvidia H100, H200, or GB300 chips

Open-source serving software vLLM announced plans to publish demonstration code showing Reflection AI's Beam model running in the 4-bit MXFP4 format supported by AMD and Nvidia Blackwell chips