DeepSeek V4.1 Flash 1017: what changed, what it costs, and how to pin or self-host it

On 17 October 2026 DeepSeek shipped DeepSeek-V4.1-Flash-1017, a new snapshot of the V4.1 Flash model it released in September, and put it behind the same API name, deepseek-flash, with no change to how you call it or what it costs (change log). If your agent or pipeline calls deepseek-flash, it has been talking to a different model since 17 October. The weights are on Hugging Face under the MIT licence the same day (model card).

This article covers what changed in the 1017 snapshot, what DeepSeek's own benchmarks claim and how far to trust them, what it costs at peak and off-peak, how to call it with thinking mode, the new Search and Fetch API that shipped alongside it, and what is and is not known about running the weights yourself. It is for engineers who already use DeepSeek or are pricing it against other models for coding and agent work.

What is DeepSeek V4.1 Flash 1017?

It is a minor update to DeepSeek-V4.1-Flash, released on 10 September 2026, with better scores on reasoning, coding, agent and visual benchmarks and, DeepSeek says, shorter outputs and faster task completion — "a Pareto improvement in performance and cost" (announcement). The architecture is unchanged from September: a 552B-parameter mixture-of-experts model on what DeepSeek calls a Causal Encoder–Decoder design, with native image understanding and a KV cache that needs a quarter of the HBM of the previous generation (V4.1 Flash announcement).

DeepSeek says the snapshot was trained with reinforcement learning from self-feedback on end-to-end project work, and is tuned for its own DeepSeek Harness (v0.2.1-rc.1 or later recommended) and for Tencent's CodeBuddy and WorkBuddy (announcement).

The deepseek-flash alias now serves a different model

Calls to deepseek-flash are now answered by the 1017 snapshot; DeepSeek's change log says to "simply use deepseek-flash to access the latest model" and that nothing about the calling method or price changed (change log). The retired names deepseek-v4-flash and deepseek-v4-flash-vision-exp are still accepted and are served by the same snapshot at the Flash price (pricing).

That is convenient, and it is also a silent model swap in production. Anything you tuned against the September model — prompts, tool schemas, output parsers, cost budgets, eval baselines — is now running against a model with different behaviour, including deliberately shorter outputs. The diagram shows where each route ends up and which ones let you choose the version.

If you need a stable version, your options are to self-host the weights, or to use a third-party router's dated model slug such as deepseek/deepseek-v4.1-flash-1017 on OpenRouter, which lists DeepInfra, inference.net, Cloudflare and DeepSeek itself as providers. Check how your router behaves when a provider retires a snapshot before relying on it.

DeepSeek V4.1 Flash 1017 benchmarks: what changed since September

On DeepSeek's own numbers, 1017 beats the September Flash on every benchmark in the model card, with the largest gains on terminal and coding-agent tasks. It also beats Claude Opus 5 on DeepSWE v1.1 (77.0 against 74.0), Automation-Bench and Apex Bench, but trails it clearly on Terminal-Bench 4.0, ProgramBench, NL2Repo-Bench and Humanity's Last Exam (model card).

Two caveats apply. Every number here is DeepSeek's, measured with harnesses DeepSeek chose: the card says Terminal-Bench 4.0 used the Claude Code harness and DeepSWE used Mini-SWE-V2. And benchmark gains do not tell you about regressions on your own tasks. Treat the chart as a reason to run your evals, not a substitute for them.

How much DeepSeek V4.1 Flash costs: peak and off-peak pricing

DeepSeek charges per million tokens with a peak/off-peak split: off-peak is half price. Peak hours are 01:00–04:00 and 06:00–10:00 UTC, Monday to Friday, excluding Chinese public holidays; everything else, including weekends, is off-peak (pricing).

Batch work that can wait for a European evening or a weekend halves its bill without any code change beyond scheduling. Note the cache-hit price: at 0.003 per million tokens off-peak, long shared system prompts and tool definitions are almost free once cached, so keep the prefix of your prompts stable. Third-party providers on [OpenRouter](https://openrouter.ai/deepseek/deepseek-v4.1-flash-1017) list lower input prices, from 0.10 per million, with different throughput and data-handling terms.

Call DeepSeek V4.1 Flash with thinking mode

The API speaks the OpenAI Chat Completions format at https://api.deepseek.com and the Anthropic Messages format at https://api.deepseek.com/anthropic. Thinking is on by default; the documented request sets it explicitly with a reasoning effort (docs):

curl https://api.deepseek.com/chat/completions \  -H "Content-Type: application/json" \  -H "Authorization: Bearer ${DEEPSEEK_API_KEY}" \  -d '{        "model": "deepseek-flash",        "messages": [          {"role": "system", "content": "You are a helpful assistant."},          {"role": "user", "content": "Hello!"}        ],        "thinking": {"type": "enabled"},        "reasoning_effort": "high",        "stream": false      }'

From Python, the official OpenAI SDK works with a different base URL. The thinking field is DeepSeek-specific, so it goes through extra_body:

import osfrom openai import OpenAI client = OpenAI(api_key=os.environ["DEEPSEEK_API_KEY"], base_url="https://api.deepseek.com") resp = client.chat.completions.create(    model="deepseek-flash",    messages=[{"role": "user", "content": "Explain the difference between a mutex and a semaphore in two sentences."}],    reasoning_effort="high",    extra_body={"thinking": {"type": "enabled"}},)print(resp.model)print(resp.choices[0].message.content)print(resp.usage)

Log resp.model and resp.usage on every call. The model field tells you which snapshot answered if DeepSeek exposes it, and the usage numbers are how you will see whether the promised shorter outputs actually lower your cost per task.

DeepSeek's new Search and Fetch API

Alongside the model, DeepSeek opened its in-house web search to developers as two endpoints, POST /v1/tools/search and POST /v1/tools/fetch, each at $0.75 per 1,000 successful requests, limited to 5 requests per second per account (Search guide). Only HTTP 200 responses with a result are billed.

curl https://api.deepseek.com/v1/tools/search \  -H "Authorization: Bearer ${DEEPSEEK_API_KEY}" \  -H "Content-Type: application/json" \  -d '{    "query": "DeepSeek V4 release",    "count": 5,    "need_content": true  }'

query takes up to 128 characters and supports operators such as site:, -site:, lang:, after: and before:; count runs from 1 to 50. Each result carries a rerank_score from 0 to 3 and an authority score from 0 to 4, which an agent can use to drop weak sources before they reach the context window. Fetch returns a page as Markdown. Search and Fetch calls made as server-side tools through DeepSeek Harness, Claude Code or Codex are free until billing starts at 00:00 Beijing time on 1 November 2026, which is 16:00 UTC on 31 October (pricing).

Can you run DeepSeek V4.1 Flash 1017 locally?

Yes, if you have a multi-GPU server; no, on a laptop. The weights are on Hugging Face in BF16/FP8 safetensors, and the card gives one-line launches for vLLM, SGLang and Docker Model Runner (model card):

pip install vllmvllm serve "deepseek-ai/DeepSeek-V4.1-Flash-1017"

DeepSeek's own inference README converts the checkpoint for 8-way model parallelism with FP4 experts (--model-parallel 8 --expert-dtype fp4) and launches with torchrun --nproc-per-node 8. It does not state the GPU model or memory required.

There is no Jinja chat template in the repository. DeepSeek ships a Python reference encoder with tests for multi-turn chat, tool calls, thinking mode and images (encoding README), and recommends its Rust deepseek-recipe libraries for production prompt building. If your serving stack assumes a chat template, check that it uses DeepSeek's encoding before you trust tool-call output.

Community quantisations appeared within a day. In the Hacker News thread, one user reported a Q4_K_S quant running "rock solid" while a 3-bit quant looped; that is a single anecdote, but it matches the usual advice to test low-bit quants on long agent runs before trusting them.

Sources and further reading

• DeepSeek: DeepSeek-V4.1-Flash-1017 announcement, API change log, pricing, first API call, Search & Fetch guide
• DeepSeek: DeepSeek-V4.1-Flash announcement (10 September 2026)
• Hugging Face: DeepSeek-V4.1-Flash-1017 model card, inference README, encoding README
• GitHub: deepseek-ai/deepseek-recipe
• OpenRouter: DeepSeek V4.1 Flash 1017 providers
• MarkTechPost: DeepSeek just released V4.1-Flash-1017
• Hacker News: model card thread and OpenRouter thread