DeepSeek V4.1 Flash 1017: MIT Weights and a Paid Search API

DeepSeek released DeepSeek-V4.1-Flash-1017 on October 17, 2026. Its model card calls it "a minor update" to V4.1 Flash that improves reasoning, coding, agent and visual results while cutting how long tasks take and how much the model writes. The weights are on Hugging Face under the MIT licence. Alongside the model, DeepSeek opened two web APIs, Search and Fetch, at $0.75 per 1,000 successful calls each, and set October 31 as the day calls made through Claude Code and Codex stop being free.

Three groups need to act on this, for different reasons. Teams already calling DeepSeek's Flash model are now on 1017 whether they asked for it or not, because the old model names point to it. Developers who run Claude Code or Codex against DeepSeek's API have web searches that will start appearing on their bill. And anyone comparing cheap models for agent work has a new data point, which we set against its rivals in our comparison of cheap agent models.

Sources for this post are DeepSeek's model card, pricing page and API documentation, all read on October 19. Every benchmark number below is DeepSeek's own.

DeepSeek treats 1017 as a dated snapshot of an existing model, not a new generation, and names it after its release day, as it did with V4 Flash 0731 in July. The base model is the one we covered at the V4.1 Flash launch in September. The model card describes the update as improving reasoning, coding, agent and visual benchmarks while reducing task execution time and output length, which it frames as a better trade between performance and cost across several agent harnesses.

The shorter-output point is the one that reaches the bill. A model billed per output token that reaches the same answer in fewer tokens is cheaper per finished task even at an unchanged price. DeepSeek does not publish a figure for how much shorter, so treat it as a direction to measure rather than a saving to budget.

The card is also blunt about what it does not include. There is no Jinja chat template, the format most inference engines expect. Instead DeepSeek ships a Python reference encoder and a set of Rust libraries with Python bindings, called deepseek-recipe, that convert standard API request formats into DeepSeek's prompt format and parse the output back. For teams calling the hosted API this changes nothing; for teams self-hosting it is the first thing to wire up.

DeepSeek's card compares 1017 with the previous V4.1 Flash and with four other models, Kimi K3, GLM-5.3, GPT-5.6 Sol and Claude Opus 5. The table below keeps the previous version and Opus 5, the strongest column on most rows, so the size of the update and the remaining gap are both visible. DeepSeek ran Terminal-Bench 4.0 in the Claude Code harness and DeepSWE v1.1 in the Mini-SWE-V2 harness.

Two readings. On DeepSeek's own agent-style rows, Automation-Bench and Agents' Last Exam, 1017 is ahead of every model in its table, including Opus 5. On the harder coding and reasoning rows, Opus 5 keeps a clear lead: nearly ten points on Terminal-Bench 4.0, seven on NL2Repo-Bench and fourteen on HLE without tools. A model priced in cents per million tokens getting this close on some rows is the news; claiming it beats a frontier model across the board is not supported by the table.

The failure case is the usual one with vendor tables: the vendor chose the benchmarks, the harnesses and the settings. DeepSeek is unusually specific about the harnesses, which helps, but no independent party has rerun these rows. Run your own agent tasks before moving a production workload, and compare against the model you use today rather than against the columns DeepSeek chose.

DeepSeek's pricing page lists the model under the name deepseek-flash, with DeepSeek-V4.1-Flash-1017 as the version behind it. A footnote adds the part most teams will miss: the legacy names deepseek-v4-flash and deepseek-v4-flash-vision-exp are still accepted, but the models behind them have been retired, and those requests are now served by 1017 and billed at the Flash price.

That is a silent change for anyone who pinned an old name to keep behaviour stable. The example: a support pipeline tuned in the summer on deepseek-v4-flash, with a parser built around that model's output style, is now talking to a different model. If 1017 writes shorter answers, as DeepSeek says it does, a parser that expected a certain structure or length may start failing quietly.

The same applies through routers. OpenRouter's ~deepseek/deepseek-flash-latest alias moved to the 1017 route after our October 16 catalogue capture. An alias exists to follow the latest version, so this is working as designed; it is still a change to log. Our model alias and retirement ledger tracks this pattern across vendors.

DeepSeek prices Flash by the clock. Off-peak, a million input tokens cost $0.15, or $0.003 when they are a cache hit, and a million output tokens cost $0.60. Peak rates are exactly double. Peak hours are 01:00–04:00 and 06:00–10:00 UTC, Monday to Friday, excluding Chinese public holidays; weekends are off-peak all day. For a team in central Europe, the second window runs 08:00 to noon until the clocks change on October 25, then 07:00 to 11:00.

The cache line is where the money is for agents. A coding agent re-reads the same files on every turn, and DeepSeek bills repeated input at a fiftieth of the normal rate. A worked example: a session that sends 2 million input tokens and receives 150,000 output tokens costs $0.39 off-peak if none of the input is cached ($0.30 plus $0.09). If 1.8 million of those input tokens are cache hits, the same session costs about $0.13 ($0.0054 for the cached input, $0.03 for the rest, $0.09 for output). Run it in a peak window and each figure doubles.

The failure case is reading a router's headline price as DeepSeek's. Since our October 16 capture, OpenRouter's headline for the older deepseek/deepseek-v4.1-flash route dropped to a few cents as provider mix shifted. That is a third-party route, not a DeepSeek price change. Quote DeepSeek's page, with the time window.

The second half of the announcement is a new product line. DeepSeek opened the search engine behind its own chatbot as two raw JSON APIs, using the same API key as its models. Its Search and Fetch guide describes Search as finding web pages for a query and returning structured results, and Fetch as retrieving a single page and returning its content as Markdown. These are the two calls a research agent or a retrieval pipeline makes most often.

A Search result carries a title, URL, short snippet, a longer summary written for a model to read, the publication date when known, an authority tier from 0 to 4 and a relevance score from 0 to 3. Queries accept familiar operators: site: to restrict to a domain, -site: to exclude one, quoted phrases, lang:, and after: and before: for dates. Up to 50 results come back per call.

For context on price, our comparison of web search APIs for agents lists what the other providers charge and limit, from $0.25 to $35 per 1,000; DeepSeek's $0.75 sits near the bottom, with one listed provider cheaper. The failure case is choosing on price alone. A search API's value is the quality and freshness of what it returns for your queries, and a China-based index may rank sources differently from the engines your users know. Run the same 50 real queries through two providers and read the results before switching.

DeepSeek's API is compatible with Anthropic's message format, and DeepSeek documents how to point Claude Code at it. When you do, Claude Code's server-side web tools run on DeepSeek's side, through these same Search and Fetch APIs. The guide says those server-side calls from DeepSeek Harness, Claude Code and Codex are currently free for API users, and that per-call billing starts at 00:00 Beijing time on November 1, which is 16:00 UTC on October 31.

The cost is easy to work out once you know your call volume. An agent that makes 40 searches and 20 page fetches per task, running 100 tasks a day, makes 6,000 calls a day. At $0.75 per 1,000 that is $4.50 a day, or about $135 over 30 days. For many teams that is small next to token spend; for a research agent that searches heavily on every step, it may not be.

The failure case is not knowing the volume. Coding agents decide for themselves when to search, and nobody counted while it was free. Use the remaining days to log how many web tool calls your sessions make, then decide whether to keep them on DeepSeek, cap them, or route search elsewhere. Our guide to running a cheaper model inside Claude Code summarises the DeepSeek setup as well.

The repository and weights are licensed under MIT, one of the most permissive licences available: commercial use, modification and redistribution, as long as the copyright and licence notice travel with the code. Hugging Face lists the model at 763 billion parameters. DeepSeek recommends temperature 1.0, top-p 1.0 and a maximum output of at least 256K tokens, with a one-million-token context.

That puts 1017 with Reflection's Beam, released a day earlier under Apache 2.0, as a large open model with no revenue trigger, and in a different position from Meta's Muse Spark Open, whose licence adds attribution and a clause for large hosting providers. We covered DeepSeek's earlier MIT release of a Flash vision build in our V4 Flash vision weights post.

Self-hosting a 763-billion-parameter model is a data-centre job, and the missing chat template adds integration work. For most teams the licence matters less for running the model than for what it permits downstream: fine-tuning, distillation into a smaller model, or serving it inside a product without a terms review. Hosted providers already list it, which is the practical route for anyone who wants the model without the hardware.

Most of what changed lands without anyone deploying anything, which is why it is worth a short checklist. The routing below is ours, from DeepSeek's documentation, not from a trial of our own.

Teams that want help moving agent workloads to cheaper models without breaking their pipelines can talk to us about our AI transformation work.