On 17 October 2026, DeepSeek quietly posted DeepSeek-V4.1-Flash-1017 to Hugging Face: a post-training refresh of September’s V4.1-Flash that is a much bigger deal than its “minor update” label suggests. The weights are out under the MIT license, the model completes agent tasks faster at the same low API price, and DeepSeek opened its in-house search engine to every developer as a paid Search & Fetch API. Below is everything you need to actually use it — the API, the download, the honest hardware math for a local run, and what the new endpoints cost.
What V4.1-Flash-1017 is — and what changed
Flash-1017 is a refreshed checkpoint of DeepSeek-V4.1-Flash, DeepSeek’s smallest model in its new architecture family. The official model card on Hugging Face calls it “a minor update” that reduces task execution time and output length, while delivering “substantial improvements across reasoning, coding, agentic tasks, and visual benchmarks.” In plain terms: same low price, faster completions, shorter (cheaper) outputs, and higher scores across the board.
Here is the release table from the model card. Treat every number as DeepSeek’s own testing — no independent lab has verified these yet:
| Benchmark | Flash-1017 | Old Flash | Opus 5 | GPT-5.6 Sol |
|---|---|---|---|---|
| DeepSWE v1.1 (coding) | 77.0 | 74.2 | 74.0 | 73.0 |
| Terminal-Bench 4.0 | 42.3 | 31.2 | 51.8 | 39.9 |
| HLE w/o tools (reasoning) | 42.6 | 36.8 | 56.3 | 44.5 |
| Automation-Bench v1.0.6 | 60.5 | 54.8 | 50.3 | 45.8 |
| Agents’ Last Exam | 35.4 | 31.8 | 28.6 | 26.7 |
Two things stand out. First, Flash-1017 beats both Anthropic’s Claude Opus 5 and GPT-5.6 Sol on DeepSWE, a software-engineering benchmark that matters for agent coding. Second, seven of the eleven benchmarks DeepSeek published still go to Opus 5 — on raw reasoning (HLE), Opus 5’s 56.3 is nearly 14 points clear of Flash-1017’s 42.6. DeepSeek is winning on agent-flavoured coding benchmarks at a fraction of the price; it is not claiming reasoning supremacy. Source: DeepSeek-V4.1-Flash-1017 model card, Hugging Face.
| Spec | Flash-1017 |
|---|---|
| Architecture | 552B-parameter MoE; 8B active parameters for input, 16B for output (official launch post, via AI Weekly) |
| Context window | 1M tokens; max output at or above 256K |
| Recommended sampling | temperature 1.0, top_p 1.0 |
| License | MIT — repository and model weights |
| API identifier | deepseek-flash (unchanged from September) |
| Chat template | No Jinja template ships; an encoding.py reference implementation plus the deepseek-recipe toolkit cover prompt encoding |
Three ways to use Flash-1017 right now
The fastest path is the API, the most flexible is the open weights, and the laziest is the app. Which one you pick depends on whether you want speed, control, or convenience.
1. The API: same ID, new brain
API customers reach Flash-1017 through the existing deepseek-flash model ID at unchanged prices: $0.15 per million uncached input tokens and $0.60 per million output tokens off-peak, with cache reads at $0.003 per million (AI Weekly, 17 Oct 2026).
Important quirk: the update shipped silently. DeepSeek replaced the model behind deepseek-flash with no changelog, so existing consumers absorbed the behaviour change — faster completions, shorter outputs — before documentation appeared. If you have prompts tuned for the September checkpoint, re-run your evaluations rather than assuming nothing changed. Source: SiliconANGLE, 18 Oct 2026.
2. Download the weights: MIT-licensed, ~510 GB
The full weights are on Hugging Face at deepseek-ai/DeepSeek-V4.1-Flash-1017 under the MIT license — free to use, modify, and commercialise. Bloomberg-origin reporting puts the release at 48 files totalling 510.3 GB. The MIT license removes the usual open-versus-closed trade-off for teams pricing a switch away from OpenAI or Anthropic.
3. The app: zero setup
If you just want to chat, the updated model is already live in DeepSeek’s web and mobile apps. No code, no bill — the simplest way to feel what “faster task execution” means before you commit to anything.
The honest hardware math for a local run
The full 510 GB weights need datacentre-class hardware; a heavily compressed copy can run on one Mac Studio — with real capability trade-offs. An independent experiment published on Medium on 19 October 2026 squeezed the model down to 114 GB and ran it on a single 128 GB Mac Studio at 35 tokens per second. On 20 recent LiveCodeBench problems, the compressed local copy solved 13; the full model via DeepSeek’s API solved 15. Source: independent Medium test, 19 Oct 2026.
The most interesting finding was not the 2-point gap. The tester initially got only 4 of 20 right, then read the outputs and discovered the compressed model was not failing to solve problems — it was failing to stop: on one problem it found the correct formula early, then spent 12,000 tokens enumerating cases until it ran out of room. A simple instruction telling it when to stop thinking recovered most of the gap, from 4/20 to 13/20.
Practical guidance, step by step:
- Check your budget for storage and bandwidth first. 510 GB of weights is the entry ticket for full-precision local hosting. That means an enterprise GPU cluster, not a gaming PC.
- For hobbyists and solo devs, the API is the rational route. At $0.15 per million input tokens, the hosted model is cheaper than any hardware you would buy to serve it.
- For the “runs on my desk” experiment, follow the compressed path. You need roughly 128 GB of unified memory (Mac Studio-class) and a quantised build, plus explicit reasoning-stop instructions in your prompts — the Medium test shows the stopping problem is real.
Caveat: this is desk research, not a hands-on test. We did not run the weights ourselves; the local-run numbers are one independent experimenter’s, published 19 October 2026.
Search & Fetch API: DeepSeek’s own search engine, now public
DeepSeek has opened the in-house search engine its agents use to every developer — and it is the second half of this release. Two raw JSON endpoints, billed separately from tokens:
- Search API — finds web pages for a query and returns up to 50 ranked results:
POST https://api.deepseek.com/v1/tools/search - Fetch API — retrieves a single page and returns its content as Markdown:
POST https://api.deepseek.com/v1/tools/fetch
Both use the same bearer API key as the DeepSeek API. Each search result carries a rerank_score (0–3; 1.5+ is high quality), an authority tier (0–4; 3+ is highly authoritative), a long model-readable summary, and optionally full page content. Queries support site:, -site:, exact-phrase, lang:, and after:/before: date operators. Source: DeepSeek API docs.
Billing is per successful request, not per token or per result: $0.75 per 1,000 successful requests on each endpoint. Failed requests, empty results and timeouts are not billed. Crucially, calls routed through DeepSeek Harness, Claude Code or Codex are free until 00:00 Beijing time on 1 November 2026 — a two-week grace period that makes this the cheapest window to experiment. Source: DeepSeek API docs, billing section.
Quick test with curl (replace the key placeholder):
curl https://api.deepseek.com/v1/tools/search \
-H "Authorization: Bearer YOUR_KEY" \
-H "Content-Type: application/json" \
-d '{"query": "DeepSeek V4.1 release", "count": 5, "need_content": true}'What it costs per task: independent arithmetic
An independent 56-trial test found Flash-1017’s per-run coding cost 3.6× below GPT-6.1 Sol on identical bounded tasks: $0.016 versus $0.058 per run. That is AI Weekly’s own hosted testing, published 17 October 2026 — not vendor marketing. The same test honestly notes that Flash-1017 trails Claude Opus 5 on 7 of 11 benchmarks. Source: AI Weekly.
Our take: who this is actually for
Flash-1017 is for agent builders who need cheap, fast, licensable intelligence — with two asterisks. The MIT weights plus a $0.75-per-thousand search API is a pricing structure aimed squarely at teams running coding agents at scale: it is why DeepSeek tuned this release for its Harness tool, Anthropic’s Claude Code and OpenAI’s Codex rather than for chat users. Pair this with our OpenCode 2.0 setup guide if you are assembling an agentic coding workflow, and our NVIDIA Nemotron 3.5 open-models guide to compare the open-model options.
The asterisks. First, reliability: days before this release, DeepSeek suffered a nearly five-hour API outage on 15 October — counted by Yicai Global among its worst this year — and some experts blamed a shortage of computing capacity. Users got credits, and DeepSeek promised to remove single points of dependency, but teams building on deepseek-flash should architect for that history. Second, the silent update itself: a model that can change under your API identifier without a changelog is a dependency you should version-pin around, not blindly ride.
The market noticed the release fast. Bloomberg reported on Saturday 18 October that shares of Chinese rivals MiniMax Group and Alibaba fell in Hong Kong trading — investors read Flash-1017 as a price war, not a press release. Source: SiliconANGLE, 18 Oct 2026.
Five questions still unanswered: will DeepSeek announce future checkpoint swaps, or stay silent? What happens to Search pricing after the 1 November free window ends? Which quantisation recipe reproduces the 114 GB Mac Studio run? How does Flash-1017 behave on non-English agent tasks? And does the MIT license cover the deepseek-recipe encoding toolkit too? We will update this guide as answers land. If you are auditing how AI tools enter your company, our Shadow AI audit framework is the companion read.
The weights are free under the MIT license — you can download, modify and commercialise them. The hosted API is pay-per-use at $0.15 per million input tokens and $0.60 per million output tokens. The Search and Fetch APIs cost $0.75 per 1,000 successful requests each, but are free through DeepSeek Harness, Claude Code or Codex until 1 November 2026.
Not in any practical sense. The full weights are about 510 GB, which needs datacentre-class hardware. An independent experiment compressed the model to 114 GB and ran it on a single 128 GB Mac Studio at 35 tokens per second — consumer laptops do not have the memory headroom.
Flash-1017 is a minor post-training update to the September V4.1-Flash checkpoint. DeepSeek says it completes tasks faster with shorter outputs, and its own release table shows gains of 2.0–11.1 points across all 11 published benchmarks. It shipped under the same deepseek-flash API identifier, so API users were moved automatically.
$0.75 per 1,000 successful requests for the Search endpoint and the same for Fetch, billed per successful request rather than per token. Failed requests and empty results are not billed. Server-side calls made through DeepSeek Harness, Claude Code or Codex are free until 00:00 Beijing time on 1 November 2026.
No. DeepSeek kept the existing deepseek-flash model identifier and unchanged prices, so your existing code calls the new checkpoint automatically. The catch is that the swap happened with no changelog, so re-run your prompt evaluations to confirm behaviour matches your expectations.
Sources and methodology
Sources: DeepSeek-V4.1-Flash-1017 model card, Hugging Face; DeepSeek Search & Fetch API documentation; SiliconANGLE, 18 Oct 2026; AI Weekly, 17–18 Oct 2026; independent local-run experiment, Medium, 19 Oct 2026. Method: desk research on primary sources; all benchmark figures are DeepSeek’s own published numbers and are labelled as such; the local-run and per-task cost figures are independent experiments, not our own testing — we did not run the model ourselves. Verified 19 October 2026.
Arva Rangwala covers AI news, AI tools, guides and prompts for OpenAIMaster — what is new, what is worth using, and how to put AI to work.