Speed in artificial intelligence has become a question of architecture, not of marketing. The debut of DeepSeek V4.1 Flash 1017 Ultrafast on Venice should be read that way: not a new model, but a faster serving route for one of the most talked-about open models of the moment.

What actually changed

On paper the family is the same: a multimodal Mixture-of-Experts with a 552-billion-parameter backbone, 8 active on input and 16 on output, and a context window of one million tokens. The Ultrafast variant changes the delivery path, tuned for agentic coding, and stays classified as private.

The price, however, is not the standard version’s: 3 dollars per million input tokens and 12 for output, against 0.33 and 1.32 for Uncensored, which the app presents as the default. Input cache costs 0.06 dollars. Anyone choosing speed pays roughly nine times more: a figure worth more than any slogan, and one worth checking before moving an agent onto this route.

Discover, without a public page

In the same hours Venice handed out early-access codes for Discover, gathering hundreds of replies. No documentation for the product exists yet: the official blog does not describe it, and the last update dates to 3 November. Early access without a public page is legitimate, but it also says that communication is moving faster than documentation.

The token, without invented numbers

The economic picture is less bright than the product. VVV trades just above 20 dollars, with a market capitalisation around 1.016 billion and a decline of roughly 23 percent over thirty days.

Inflated reconstructions are circulating around the numbers: better to pin down the verifiable ones. DefiLlama books a TVL of zero dollars by its stated methodology and keeps staking as a separate series, 728.39 million dollars, 71.69 percent of market capitalisation. Cumulative fees, which by methodology equal the dollar value of VVV bought back and burned, are 4.11 million.

The official tracker reports 33.9 million VVV burned, 41.8 percent of the reference supply, 25.2 million staked, 8.4 million locked and 15.5 million circulating. In the last day the burn rate was 2.72 thousand VVV for 60.05 thousand dollars, against 63.14 thousand the previous day and an average of 55 thousand. Since 5 November, moreover, Valour lists a VVV exchange-traded product in Stockholm denominated in Swedish kronor, with a 1.9 percent fee.

The point

Venice is doing two things at once: selling speed to whoever pays for it, and opening channels to the token for whoever does not want a wallet. The moves are coherent, but they do not reinforce each other: a premium model does not move market capitalisation, an exchange-traded product does not make inference faster. The question is whether that million a month will accelerate, or stay the accounting shadow of a product growing somewhere else.

Verified sources: DeepSeek V4.1 Flash 1017 Ultrafast model page on docs.venice.ai and the model index (275 families, 397 IDs, updated 6 November 2026); @AskVenice post of 6 November 2026 on the model and the Discover codes; official venice.ai changelog, consulted 6 November 2026; defillama.com/protocol/venice (TVL 0 $, staking 728.39 mln, cumulative fees 4.11 mln, price 20.78 $, mcap 1.016 bln, consulted 6 November 2026); official tracker venice.ai/token/burns read on 6 November 2026 (33.9 mln burned / 41.8%, 25.2 mln staked / 31.1%, 8.4 mln locked, 15.5 mln circulating); GlobeNewswire press release of 5 November 2026 as carried by Placera, stocktitan and CoinMarketCap (VVV ETP, ISIN CH1108681771, 1.9% fee). Discrepancies noted: Echo’s report indicated a TVL of 1,456,774,920 dollars, which does not exist on DefiLlama, and an average daily burn spend of 29.4 thousand dollars with a positive sign, while the tracker shows a 55 thousand average and -5% on the previous period.