The Day in AI

• DeepSeek released V4.1-Flash-1017 under the MIT license and reports 42.6 on Humanity's Last Exam without tools, up from 36.8.
• DeepSeek also added standalone Search and Fetch endpoints to its API, each priced at 5 yuan per 1,000 successful requests.
• Tencent Hunyuan open-sourced Hy Memory 2.0, an agent memory stack that adds working memory and code memory to its chat memory.
• Shanghai AI Laboratory released the full version of Atria Dawn, a 744B agent model with a 512K context window, through hosted APIs only.
• Elorian released Glimpse, its first model, which crops images and pulls video frames while it reasons, at $0.60 per million input tokens.
• Replit is no longer working with Anthropic, its CEO told TechCrunch.

DeepSeek calls V4.1-Flash-1017 a minor update to the V4.1-Flash it released in early September. Its technical report claims gains in reasoning, coding, agentic tasks and visual benchmarks, and says the model finishes tasks faster with shorter outputs. On the text-only subset of Humanity's Last Exam it reports 44.0, up from 39.1. All of these figures are DeepSeek's own. The model is live on the web, in the app and on the developer platform, with weights on Hugging Face.

On the same day DeepSeek added two web endpoints to its API. Search queries the web and returns 1 to 50 structured results, each with a snippet, a summary written for a model to read, a rerank score and an authority rating. Fetch retrieves one URL and returns its title and body in Markdown. Each costs 5 yuan per 1,000 successful requests and is capped at 5 queries per second per account, and failed or empty calls are free. An agent built on DeepSeek can now get live web content without a separate search provider. Search calls from Claude Code's Web Search tool or Codex's web_search are free for the first two weeks.

Tencent Hunyuan's Hy Memory 2.0 has an SDK, a memory model called Hy-Memory-30B-A3B and a local app called MemHub, all open-sourced and usable separately. The new working and code memory record a task's goal, decisions, intermediate results and lessons. Before it retrieves any of them, the system works out which task is in progress. Tencent reports 72.6% on LoCoMo for chat memory, with tokens per question cut from 1,159 to 703. It also reports a macro average of 55.49 for its memory model on a 20% subsample of benchmarks, against 54.62 for Claude Opus 5.

Shanghai AI Laboratory's Atria Dawn is a 744B mixture-of-experts model built on a GLM-5.2 base, and the full version follows a September preview. The lab reports 57.5 on JobBench against 50.3 for the preview, and 61.7 on SWE-bench Pro. The preview still scored higher on BrowseComp, with 92.5 against 91.3. The GitHub download table still lists a 256K context, and the full weights are listed as coming soon.

• Black Forest Labs released FLUX 3 Image Klein on 15 October, a 5.6B open weights image model under a restricted license whose INT8 build runs on RTX GPUs with 8GB of memory.
• Elorian released Glimpse on 16 October as a research preview that returns to the image mid-reasoning to crop regions, pull video frames or run code, priced at $0.60 per million input tokens and $3 per million output tokens.
• NVIDIA released Nemotron-Labs Dory1.0-4B on 16 October, a 4B text model with a thinking mode, and says its NVFP4 build runs 2.17 to 2.44 times faster than BF16 on DGX Spark.
• Replit is no longer working with Anthropic, its CEO told TechCrunch.
• Meta removed terms of service language that suggested it could scrape websites associated with business accounts, saying it applied only to ads and business products, The Verge reported.