SShortSingh.
Back to feed

DeepSeek V4.1-Flash cuts KV cache 4x despite doubling model size, targeting AI agents

0
·72 views

DeepSeek released its V4.1-Flash model on September 10, 2026, specifically engineered to reduce the cost of running AI agents that repeatedly read large contexts rather than generate long outputs. The model introduces a Causal Encoder-Decoder (CED) architecture that splits its 552 billion parameters into encoder and decoder halves, sharply reducing active parameters during input processing to just 8 billion. Additional techniques — including Compressed Sparse Attention 2 (CSA2) and FP4 KV caching — bring total KV cache storage down to 890 bytes per token, roughly one-quarter of its predecessor V4-Flash and 437 times smaller than the original V1. This matters economically because agentic workloads spend most compute re-reading the same context in loops, meaning KV cache — not token generation — is the dominant cost driver. Benchmark scores for agent tasks such as DeepSWE and Terminal-Bench rose substantially, from 54.4 to 74.2 and 82.7 to 90.6 respectively, compared to V4-Flash.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

Log in to join the discussion and vote.

Log in

Related stories

0
ProgrammingHacker News ·

Onlook Launches Open-Source Desktop Tool for Designing Directly in Code

Daniel, founder of Onlook, a Y Combinator W25 company, has launched an open-source desktop app that places a design canvas over a real codebase. The tool is aimed at designers who now need to work in code, letting them explore many ideas at once. Each design variation appears as a live frame running the actual application with real data, and changes arrive as pull requests that can be reviewed or discarded. The team said it rebuilt the product as an agentic design tool after earlier web and Storybook-based attempts, and that the project has drawn over 25,000 supporters and 100 contributors.

0
ProgrammingDEV Community ·

Running a self-built Google scraper at scale: blocks, silent failures and hidden engineering costs

A test estimate puts a self-run Google scraper at roughly $2,450 to $5,150 a month for 100,000 searches, with engineering hours making up most of that cost. The same volume through four commercial SERP APIs is priced between $200 and $725 a month. In a November 2026 test, a default headless browser was redirected to Google's /sorry/ block page on every request, and some newer failures return HTTP 200 with empty or wrong data. Google has since added JavaScript verification pages and poisoned result pages, and it stopped honoring the num=100 parameter, multiplying the requests needed per keyword. The authors argue that outsourcing to a SERP API shifts the detection burden to the vendor rather than removing it.

0
ProgrammingDEV Community ·

SerpApi Ranks Fastest Among Five Google SERP APIs in 30-Day Test

A 30-day comparison from October 20 to November 18, 2026 tested five Google SERP APIs on the same 100 short keywords daily. SerpApi was the quickest, with a median response time of 2.1 seconds and a 90th percentile of 4.9 seconds. DataForSEO, SearchApi and SerpApi had success rates of 99.8% to 99.9%, while Oxylabs failed most often at 92.4%. Failures tended to cluster on a few bad days rather than occurring at a steady rate. The results cover only whether a response arrived without error, not whether its contents were accurate.

0
ProgrammingDEV Community ·

x402 pay-per-call market in November 2026: sellers, conformance and buyers

Blockchain IQ measured the x402 market, where software pays for API calls with stablecoins, as of November 2026. On 16 November 2026, its daily roster counted 3,048 graded mainnet seller wallets listing 48,036 endpoints. Around 63% of these sellers earned an A grade in an unauthenticated probe checking whether endpoints correctly return a 402 response, while roughly one in five earned an F; no payments were sent during testing. In the week of 2 to 8 November 2026, 6,809 wallets made about 1.18 million payments worth roughly $33,281 to listed Base sellers, averaging about three cents each. The analysis relies on listed sellers only, covers a single week of buying, and infers payments from on-chain logs rather than invoices.