DeepSeek V4.1-Flash Slashes Agent Memory Costs to One-Quarter

DeepSeek V4.1-Flash Cuts Agent Memory Costs Fourfold

DeepSeek released V4.1-Flash on September 10, 2026, a 552-billion-parameter multimodal model built on four interlocking architectural techniques that together reduce the active key-value cache memory required for long-running AI agents to one-quarter of what the prior V4-Flash generation needed. The release shipped with a 50-page technical report on Hugging Face and marks the debut of the V4.1 model series, a distinct architecture lineage rather than an incremental update. Starting September 14, 2026, all API traffic to the deepseek-v4-pro endpoint is automatically rerouted to V4.1-Flash at V4.1-Flash rates, making the transition mandatory for every developer calling that endpoint. The memory savings come from a Causal Encoder-Decoder split that reorganizes the model’s 40 Transformer layers into a 20-layer causal encoder and a 20-layer decoder, Compressed Sparse Attention 2 with three cache-sharing modes, FP4 cache quantization, and the elimination of persistent sliding-window attention storage. V4.1-Flash achieves a global KV cache of 890 bytes per token, roughly one-quarter of V4-Flash and approximately 1/437th of DeepSeek-V1’s per-token global KV cache size from two years earlier. Persistent SSD storage for the KV cache drops to approximately one-eighth of V4-Flash’s requirement. The model uses Mixture-of-Experts routing with one shared expert and 384 routed experts per block, activating only six routed experts per token, an activation rate of roughly 1.6 percent. It trained on 45 trillion tokens, with the context window extended from 64K to 1 million tokens, and includes a 196-billion-parameter conditional memory module called Engram plus native multimodal support through a 32-layer vision encoder. At maximum effort, V4.1-Flash posted the highest reported scores in its comparison group on several benchmarks, including 90.6 percent on Terminal-Bench 2.1, 74.2 percent on DeepSWE v1.1, 88.1 percent on CyberGym, and a Codeforces rating of 3,471, above V4-Pro’s 3,348. DeepSeek acknowledges that larger closed models retain an advantage on specialist-knowledge-intensive tasks and notes a persistent multimodal gap relative to the largest proprietary systems. The technical report also surfaces a finding governing interpretation of agent benchmark tables: the same V4.1-Flash checkpoint scored 65.5 percent to 74.2 percent on DeepSWE v1.1 depending solely on which agent harness wrapped it, a range of 8.7 percentage points produced by changing the evaluation framework rather than the model. DeepSeek adjusted pricing for the V4.1-Flash API tier, split into peak and off-peak periods with off-peak rates at half the peak rate.

DeepSeek V4.1-Flash Cuts Agent Memory Costs Fourfold With New Architecture →

OpenAI Opens GPT-Live-1 Full-Duplex Speech Model to Developers

OpenAI has made GPT-Live-1 available to developers as an API. The speech model can listen and talk at the same time, a feature known as full-duplex, and is already running inside ChatGPT. Developers can pair it with different backend models depending on the task, matching reasoning depth, speed, and cost to each use case. It costs $0.05 per minute. Yelp is using the model for phone-based reservations and reports better call handling, according to CTO Alex Levy. On OpenAI’s benchmarks, GPT-Live-1 scores 80.1 percent in full-duplex interactivity tests, compared to 45.4 percent for GPT-Realtime-2.1. Turn-taking latency drops to 0.8 seconds from 1.4 seconds, tool-calling accuracy rises to 87 percent from 60 percent, and in a banking voice support benchmark the model reaches a 32 percent pass rate, up from 12.4 percent for the previous model. GPT-Live-1 ships with twelve new voices spanning different accents, dialects, and languages, and provides ASR transcripts and response text out of the box. Full details will be available in the API documentation.

OpenAI’s GPT-Live-1 API lets developers build apps that talk and listen at the same time →

Clearview AI Tests InquiryIQ Tool to Map People’s Online Lives for Police

Clearview AI is testing a previously unreported prototype called InquiryIQ that would let law enforcement surface associates, social accounts, and other information about people identified through Clearview. The prototype tested a model from xAI, the maker of Grok. Clearview AI became known in 2020 for scraping more than 3 billion photos from the internet to turn faces into names for police and security professionals. The new tool is intended to help law enforcement fill in details about the person behind a name, including who they are, who they know, where they live, and what they have left behind online.

Clearview AI Is Testing an AI Tool That Would Let Cops Unearth Your Life Online →

MacBook Uses Mirror and Webcam to Let AI Agent Debug Its Own GPU Drivers

Linux developer Justin Schroeder shared a photo of a laptop using realtime visual feedback during an AMD Radeon GPU driver tuning task. Schroeder said the MacBook is using its webcam to look at its screen in a mirror to improve AMD Radeon chip support in Omarchy. The machine is presumed to be running a programming agent such as Claude Code that watches its own screen to assess the GPU driver tweaks it makes. Because the MacBook is working on itself, it must be an older Intel Mac with an AMD GPU inside, allowing the coding agent to refine Radeon hardware support and benefit from the webcam’s visual feedback. Epic Games boss Tim Sweeney commented humorously on the use of AI, saying it gave him HAL 9000 lip-reading vibes. Omarchy is a Linux distribution tailored for the age of agents, with specialized drivers and configurations suited to users of older Intel-based Macs. Its homepage advertises lightning-fast installation and built-in agents that can debug issues, letting users vibe their way through every alteration, tweak, or trouble. Omarchy is also available for Apple Silicon Macs and modern x86 PCs, and is suitable for low-spec machines such as a 2011 ThinkPad X220 with 2GB of RAM, according to the developers. It is distributed under the MIT license.

Old MacBook uses a mirror, webcam, and AI agent to code its own AMD GPU drivers — ‘agent-first’ Omarchy Linux debugs itself, AI can check its own progress on screen in real-time →