DeepSeek V4.1 Flash 1017: what's new, parameters, price and VRAM requirements

Update, October 20, 2026: A DHS order filed October 19, 2026 bars DeepSeek products, including third-party products built on DeepSeek, from US federal civilian agencies. See DeepSeek banned in US federal agencies.

DeepSeek released DeepSeek V4.1 Flash 1017 on Saturday, October 17, 2026, a new checkpoint of the open-weight V4.1-Flash model it shipped on September 10. DeepSeek's model card calls it "a minor update." It is the same architecture and size, with better scores on DeepSeek's own benchmarks and shorter outputs. It keeps the MIT license and the API price. The same day, DeepSeek opened a separate, paid Search and Fetch API. Some coverage has described the release as a model with "live web search," which isn't quite right; details below.

This story is based on DeepSeek's Hugging Face model card and config, DeepSeek's API change log, pricing page and Search guide, OpenRouter's model listing, and coverage by The Decoder and Investing.com, checked October 18, 2026. DeepSeek's release note page (news261017) didn't render its content when we checked.

The short version

• Released: October 17, 2026, on Hugging Face and in DeepSeek's API (and in its website and app, Investing.com reported, citing DeepSeek's posts on X), where deepseek-flash is "upgraded to DeepSeek-V4.1-Flash-1017" with "call method and pricing unchanged," per DeepSeek's change log.
• What's new: DeepSeek reports higher scores on all 11 benchmarks in its table, plus shorter outputs and faster task completion. It hasn't published architecture changes for 1017.
• Size: 763B parameters in the published weights. The "552B" figure is DeepSeek's "backbone" count for V4.1-Flash, not the total.
• License: MIT, for code and weights.
• Price: unchanged at $0.15 input and $0.60 output per million tokens off-peak.
• VRAM: no official requirement. The repository holds 510 GB of files.

What's new in DeepSeek V4.1 Flash 1017?

DeepSeek's model card says the update improves reasoning, coding, agentic tasks and visual benchmarks while it "reduces task execution time and output length." Investing.com, citing DeepSeek's posts on X, said DeepSeek called it a "Pareto improvement" in performance and cost. The Decoder reported that DeepSeek trained it with what it calls reinforcement learning with self-feedback, in which the model's own judgments serve as the reward signal. That is DeepSeek's description; no technical report for 1017 has been published. The card still links the V4.1 technical report from September.

DeepSeek's change log says the model is tuned for its own agent harness, DeepSeek Harness (version 0.2.1-rc.1 or later recommended), and The Decoder and Investing.com add Tencent's CodeBuddy and WorkBuddy.

For the architecture itself, the causal encoder-decoder design, native vision and 1M-token context, see our September story on V4.1-Flash. The 1017 config file shows the same layout: 40 layers, 384 routed experts plus one shared expert, and 6 experts active per token.

Does DeepSeek V4.1 Flash 1017 have live web search?

Not built into the model. What DeepSeek launched on October 17 is a pair of separate tool APIs. Its Search guide describes a Search endpoint, which DeepSeek's docs say is "powered by our in-house search engine," and a Fetch endpoint that returns a page as Markdown. Developers call them like any other API:

• Price: $0.75 per 1,000 successful requests for each endpoint. Empty results, failures and timeouts aren't billed.
• Limits: 5 queries per second per account for each, as of October 18. DeepSeek's guide, checked October 20, now lists 100 per second for Search and 25 for Fetch, plus a per-domain limit on Fetch (see our web search API comparison).
• Free period: server-side tool calls made through DeepSeek Harness, Claude Code or Codex are free until per-request billing starts at 00:00 Beijing time on November 1, 2026.

How many parameters does DeepSeek V4.1 Flash have?

The Hugging Face safetensors metadata for 1017 lists 763B parameters, the same as the September V4.1-Flash. That model's card describes "552B backbone parameters" plus a 196B "Engram conditional memory," with about 8B parameters active per token while reading input and 16B while generating. The 1017 card doesn't restate those figures.

What are the VRAM and hardware requirements?

DeepSeek doesn't publish any. Neither the 1017 model card nor the V4.1-Flash card lists minimum GPUs or memory. Here's what the repository itself shows:

• The repository's files total 510 GB on Hugging Face, split across 48 safetensors shards. The September V4.1-Flash repository is also listed at 510 GB.
• The config stores weights in FP8, with experts in FP4, which is why 763B parameters take less than 763 GB.
• Those 510 GB are file size, not a requirement. A server also needs memory for the KV cache and activations, which grow with context length.

DeepSeek documents serving through vLLM, SGLang, Transformers and Docker Model Runner. Like V4.1-Flash, 1017 ships without a Jinja chat template, so your serving stack has to support DeepSeek's reference encoder. On Hacker News (Tier 3, one user's report), a commenter said a Q4_K_S community quantization was "rock solid" while a smaller Q3_XS one "would loop like crazy." For the memory math behind smaller models, see how to choose a small AI model for your laptop.

How much does the DeepSeek V4.1 Flash API cost?

Peak hours are 01:00–04:00 and 06:00–10:00 UTC on weekdays, excluding Chinese public holidays. Context is 1M tokens and maximum output 384K. The legacy IDs deepseek-v4-flash and deepseek-v4-flash-vision-exp are also served by 1017 at the Flash price.

Last verified: October 18, 2026 (DeepSeek pricing page)

OpenRouter lists the model, created October 17, from $0.10 input and $0.50 output through third-party hosts. Those are the hosts' prices, not DeepSeek's. For how per-token prices translate into real costs, see frontier APIs vs open models.

Is there a DeepSeek V4.1 Pro?

Not yet. DeepSeek hasn't released or dated a V4.1-Pro. Its pricing page now lists deepseek-v4-pro as served by DeepSeek-V4-Pro-0813 at $0.66 input and $1.98 output per million tokens off-peak, without vision. DeepSeek's change log says V4-Pro API service continues after September 14 "with billing unchanged."

How does 1017 score? (vendor-reported)

All numbers come from DeepSeek's model card, including those for the other models. DeepSeek says Terminal-Bench 4.0 ran in the Claude Code harness and DeepSWE in the Mini-SWE-V2 harness; it doesn't say how the competitor scores were produced.

*Text-only HLE subset; on that subset DeepSeek lists 1017 at 44.0. More rows are in the data file for this story.

Last verified: October 18, 2026 (DeepSeek model card; vendor-reported)

Two caveats. The US comparison columns are GPT-5.6 Sol and Claude Opus 5, which OpenAI and Anthropic replaced on September 22, 2026, so the table says nothing about GPT-6 or Opus 5.5. And in DeepSeek's own table, Claude Opus 5 still leads on Terminal-Bench 4.0, ProgramBench and HLE, among others. The Decoder, reading a DeepSeek chart, reported about $0.20 per DeepSWE task for 1017 versus about $0.50 for its predecessor; those are DeepSeek's figures too. No independent evaluation had been published as of October 18. Our guide to coding-agent evaluations explains why harness choice changes these scores.

Why it matters

DeepSeek is updating its open model in place, on a roughly monthly rhythm, without raising prices. It's doing so while reportedly raising a round of about 100 billion yuan (our funding story). The MIT license remains among the most permissive terms in our license comparison. The Search API is DeepSeek's first paid product beyond tokens, and it competes with search tools that agent builders currently buy elsewhere.

What's confirmed, reported and uncertain

• Confirmed (DeepSeek docs, Hugging Face): the October 17 change log entry, the "minor update" wording, the MIT license, 763B parameters in metadata, the 510 GB repository, unchanged API prices, the Search and Fetch API terms, and the benchmark table.
• Reported (The Decoder, Investing.com, citing DeepSeek posts): the "Pareto improvement" framing, the self-feedback training method, and the per-task cost chart.
• Uncertain: what changed in training or architecture, independent results, real serving memory needs, and when V4.1-Pro arrives.