DeepSeek V4.1 Flash: specs, price, benchmarks, and where your prompts go

The Morse teamPublished November 2, 20268 min read

DeepSeek V4.1 Flash is DeepSeek's current small open-weight model, released on September 10, 2026 and updated on October 17, 2026 as DeepSeek-V4.1-Flash-1017: a Mixture of Experts model with 552 billion backbone parameters, 8 billion active while it reads your prompt and 16 billion while it writes, a one-million-token context, image input, and an MIT license. On DeepSeek's API it's called deepseek-flash and costs $0.30 per million input tokens and $1.20 per million output tokens at peak hours, half that off-peak.

We read DeepSeek's change log, its release post and 1017 update post, its price list, its thinking-mode guide, both model cards on Hugging Face (V4.1 Flash and 1017), its privacy policy, and its open platform terms on October 24, 2026. Every benchmark score below is DeepSeek's own. DeepSeek deserves credit for publishing the full weights of each version under MIT, and for printing scores where its model loses.

Is it DeepSeek V4 Flash or V4.1 Flash?

V4.1 Flash. DeepSeek V4 Flash was the earlier model, and DeepSeek has retired it. Its change log entry for September 10, 2026 says V4 Flash and V4 Flash Vision Exp "are retired," and that requests to the old names deepseek-v4-flash and deepseek-v4-flash-vision-exp are temporarily routed to V4.1 Flash. As of October 24, 2026, DeepSeek's price list says those requests are served by DeepSeek-V4.1-Flash-1017 and billed at the Flash price.

Many of the pages ranking for "deepseek v4 flash" still describe the old model: 284 billion parameters with 13 billion active, the figures on the V4 Flash model card, which still calls itself "a preview version." If you call deepseek-v4-flash on DeepSeek's API today, a different, larger model answers. Several API resellers also list deepseek-v4.1-flash as DeepSeek's model ID. DeepSeek's own pages use deepseek-flash, and the change log says to "simply use deepseek-flash to access the latest model."

The Flash line, from DeepSeek's change log

Sources: DeepSeek's change log, read October 24, 2026. The weights for each V4.1 version went up on Hugging Face the same day as the release: the V4.1 Flash repository was created on September 10 and the 1017 repository on October 17, by Hugging Face's own records.

DeepSeek V4.1 Flash specs

The release post calls V4.1 Flash "the smallest model in our new architecture family," and names a V4.1 Pro still to come. It said deepseek-v4-pro would route to V4.1 Flash "until V4.1-Pro launches," but the change log later says V4 Pro service continues after September 14, 2026 with billing unchanged, and the price list still serves deepseek-v4-pro as V4-Pro-0813.

How good is DeepSeek V4.1 Flash?

On DeepSeek's own agent and coding tests, the small V4.1 Flash beats the bigger V4 Pro, and the 1017 update pulls ahead of the closed model DeepSeek labels "Opus 5" on several of them. These figures come from the two model cards, at maximum reasoning effort:

Sources: V4.1 Flash model card and 1017 model card, read October 24, 2026.

Two caveats from DeepSeek's own pages. The harness changes the score: on the V4.1 card, the same model gets 74.2 on DeepSWE in mini-SWE and 69.8 in Claude Code, and the 1017 card ran Terminal-Bench 4.0 in Claude Code and DeepSWE in Mini-SWE-V2. And the two cards don't always agree on the same model: NL2Repo-Bench for the original V4.1 Flash is 64.0 on its own card and 65.4 on the 1017 card. On general knowledge, the gap to the top closed model stays wide, with HLE at 42.6 against 56.3.

How much does DeepSeek V4.1 Flash cost?

As of October 24, 2026, DeepSeek's price list charges, per million tokens:

Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday, except Chinese public holidays. Every other hour, including weekends, is off-peak. That split explains why reseller listings quote two different prices for the same model: $0.30 and $1.20 is DeepSeek's peak rate, and $0.15 and $0.60 is its off-peak rate. V4 Pro costs $1.32 in and $3.96 out at peak, more than four times as much for input. The Flash API allows 2,500 concurrent requests to V4 Pro's 500.

DeepSeek's new Search and Fetch APIs cost $0.75 per 1,000 successful calls each. The 1017 post says calls made through server-side tools in DeepSeek Harness, Claude Code, and Codex were free until 16:00 UTC on October 31, 2026.

Can you run DeepSeek V4.1 Flash locally?

Only with server-sized memory. The 1017 weights come in 48 files that add up to about 510 GB, by our sum of the sizes Hugging Face lists. DeepSeek's card gives commands for Transformers, vLLM, SGLang, and Docker Model Runner, and points to community quantized builds for llama.cpp, Ollama, and LM Studio. It ships no chat template: you encode prompts with DeepSeek's reference encoding.py or its deepseek-recipe library. DeepSeek recommends a temperature of 1.0 and a max_tokens of at least 256K. DeepSeek publishes no minimum hardware.

Where do your DeepSeek V4.1 Flash prompts go?

That depends on who runs the model. The 1017 post says the model is "now live on our website, app, and API," so chat.deepseek.com, DeepSeek's app, and deepseek-flash all send your prompt to DeepSeek.

The privacy policy also lists a right to opt out of your data being used to train DeepSeek's models, by writing to privacy@deepseek.com. We read the app's policy clause by clause, including training, retention, and the government actions against DeepSeek, in is DeepSeek safe. Because the weights are MIT-licensed, the same V4.1 Flash runs anywhere, and DeepSeek's policy covers only the copies DeepSeek runs.

How to use DeepSeek V4.1 Flash without sending it to DeepSeek

Every privacy question above starts with where the prompt goes. On chat.deepseek.com and the API, it goes to servers in China, tied to a DeepSeek account. On Morse, DeepSeek V4.1 Flash is one of the unlimited text models on Pro, and you use it without a DeepSeek account or DeepSeek's app.

• Private, for everyday V4.1 Flash chats. Private is the default for a new chat, and open-weight models like DeepSeek, Kimi, and GLM run in it. Morse has zero data retention agreements with its inference, search, and other providers, so they keep nothing. When a provider confirms zero retention on a reply, Morse checks it and refuses a reply that doesn't; otherwise it relies on the provider's commitment. Here's what zero data retention covers.
• E2EE, for prompts you wouldn't send to any host. Open-weight models like DeepSeek, GLM, and Kimi also run in secure enclaves, which third-party providers operate, not Morse. Your browser checks the enclave's hardware attestation against versions pinned in the app, then seals the message so only that enclave can read it, and Morse's servers relay ciphertext they can't read. The guarantee rests on the enclave hardware and attestation working as designed. Tools in E2EE run only on your device, so switch to Private when you want web search. Our guide walks through each enclave check.
• Anonymous, to check DeepSeek's Opus comparison. DeepSeek's table puts V4.1 Flash ahead of Opus 5 on some tests and behind on others. Put the same question to Claude, GPT, or Gemini in Anonymous mode: Morse's proxy strips your account, email, and IP address, so the provider answers without knowing it's you, though it may keep the message for a limited time to check for abuse, under its own policy.

Your chat history stays in your browser, encrypted, and locked when you sign out. Morse doesn't train AI on your chats, sell your data, or show ads (privacy policy), and because it doesn't store chats, it usually has very little to give authorities. Clearing your browser deletes the history for good, so keep a copy of anything you need. For other open models on Morse, see GLM 5.3 Flash and Kimi K3.

Start free, then choose Pro at $19 a month for unlimited DeepSeek V4.1 Flash and end-to-end encrypted chats.