[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"branding":3,"analytics":7,"article-deepseek-v41-flash-cuts-costs-keeps-its-habit-of-guessing":10,"sections":35},{"siteName":4,"siteTagline":5,"publisherName":4,"contactEmail":6},"The Revision","Tech news, decoded.","editor@therevision.news",{"gaMeasurementId":8,"adsenseClientId":9},"G-ZW2MV82GYR","ca-pub-8533917693782264",{"article":11},{"id":12,"slug":13,"title":14,"dek":15,"body_md":16,"tags_json":17,"published_at":18,"created_at":19,"updated_at":20,"status":21,"review_note":22,"review_notes":23,"image_url":22,"persona_id":22,"persona_name":22,"section":24,"tags":25,"sources":30,"feedback":34,"feedback_at":22,"cost_usd":34,"total_tokens":34},12272,"deepseek-v41-flash-cuts-costs-keeps-its-habit-of-guessing","DeepSeek V4.1 Flash Cuts Costs, Keeps Its Habit of Guessing","A cheap open-weight model scores well on coding-style tasks, but a third-party test shows it rarely admits when it does not know the answer.","DeepSeek has updated its V4.1 Flash model to revision 1017, and the pitch is a sharp drop in running costs. The catch is a weak grip on what it does not know.\n\nThe model is a Mixture-of-Experts design, meaning only part of it works on any request: 16B of its 552B parameters. It accepts text and images and has a 1 million token context window. The bigger change is memory. Compared with V4 Flash, V4.1 uses one-quarter of the HBM (the fast memory on AI chips) and one-eighth of the SSD storage for its KV cache, the running record of a conversation that saves the model from recomputing it. Each token entry takes 890 bytes.\n\nAt DeepSeek's own off-peak prices, input costs $0.15 per million tokens and output $0.6, dropping to $0.003 for cache hits. Search runs $0.75 per 1,000 queries, against $1 at the nearest rival, Parallel. The weights are open, so outside hosts can undercut those numbers.\n\nDeepSeek said its internal benchmarks place the model between GPT-5.6 Sol and Claude Opus 5, beating the latter on task tests. Independent data is more measured. Artificial Analysis ranks the revision second among large open-weight models with a score of 45, well behind the 58 of Claude Opus 5.5. Only Xiaomi's MiMo V2.6-Pro beats it on cost per task.\n\nThe weakness shows up in Artificial Analysis' Omniscience test. DeepSeek answered 43% of questions correctly, strong for an open-weight model but behind proprietary ones, and it hallucinated 67% of non-answers. Because that index rewards right answers, penalizes wrong ones and ignores abstentions, the model scores 5. Kimi K3-max gets 20, and frontier models reach 42 or higher.\n\nThat shapes where it belongs. Programming, OCR with review and format conversion all give an agent something to check its work against, so mistakes get caught and retried. Open-ended research, fact lookups and precise drafting do not. Kimi K3 drew similar praise for cheap near-frontier performance and showed the same trade-off, so this looks like a pattern in the category, not a one-off.\n\nA cheap model paired with a verifier is a sound design. A cheap model left alone to report facts is not. The open question is whether open-weight labs can close the reliability gap before frontier labs raise prices, which is when these models matter most.","[\"deepseek\",\"open-weight\",\"llm\",\"benchmarks\"]","2026-10-23T12:00:00.000Z","2026-10-23T13:39:36.921Z","2026-10-23T13:39:37.121Z","published",null,[],"ai",[26,27,28,29],"deepseek","open-weight","llm","benchmarks",[31],{"name":32,"url":33},"Tom's Hardware","https:\u002F\u002Fwww.tomshardware.com\u002Ftech-industry\u002Fartificial-intelligence\u002Fdeepseeks-latest-ai-model-is-smarter-than-ever-but-its-low-running-costs-come-at-a-price-open-weight-model-is-fit-for-agentic-workloads-but-is-more-prone-to-hallucination-than-proprietary-competitors",0,{"sections":36},[37,41,46,51,56,61,66,71,76,81,86,91,96,101],{"name":38,"slug":24,"count":39,"latest_published_at":40},"AI",7072,"2026-10-23T14:18:39.000Z",{"name":42,"slug":43,"count":44,"latest_published_at":45},"Security","security",1060,"2026-10-23T11:00:00.000Z",{"name":47,"slug":48,"count":49,"latest_published_at":50},"Policy","policy",578,"2026-10-23T14:47:41.000Z",{"name":52,"slug":53,"count":54,"latest_published_at":55},"Deals","deals",569,"2026-10-23T14:43:38.000Z",{"name":57,"slug":58,"count":59,"latest_published_at":60},"Hardware","hardware",291,"2026-10-23T13:30:00.000Z",{"name":62,"slug":63,"count":64,"latest_published_at":65},"Consumer Tech","consumer-tech",290,"2026-10-23T08:23:42.000Z",{"name":67,"slug":68,"count":69,"latest_published_at":70},"Science","science",216,"2026-10-22T22:12:19.000Z",{"name":72,"slug":73,"count":74,"latest_published_at":75},"Startups","startups",147,"2026-10-23T12:44:27.000Z",{"name":77,"slug":78,"count":79,"latest_published_at":80},"Software","software",142,"2026-10-23T00:09:58.000Z",{"name":82,"slug":83,"count":84,"latest_published_at":85},"Dev Tools","dev-tools",122,"2026-10-21T16:00:00.000Z",{"name":87,"slug":88,"count":89,"latest_published_at":90},"General","general",100,"2026-10-23T13:36:50.000Z",{"name":92,"slug":93,"count":94,"latest_published_at":95},"Gaming","gaming",88,"2026-10-22T22:25:33.000Z",{"name":97,"slug":98,"count":99,"latest_published_at":100},"Reviews","reviews",38,"2026-10-20T12:00:00.000Z",{"name":102,"slug":103,"count":104,"latest_published_at":105},"How-To","how-to",8,"2026-10-05T09:00:00.000Z"]