This build suits developers who want a low-rate, self-hostable model for coding agents and need image input. It still trails Claude Opus 5 in DeepSeek's own table on Terminal-Bench 4.0 (42.3 against 51.8) and Humanity's Last Exam (42.6 against 56.3), so keep a stronger model for the hardest reasoning.
What changed
DeepSeek-V4.1-Flash-1017 is DeepSeek's 17 October 2026 open-weights update to V4.1-Flash, a mixture-of-experts model that reads images and text across a 1M-token context. DeepSeek reports 77.0 Pass@1 on DeepSWE v1.1 using the Mini-SWE-V2 scaffold, ahead of the 74.0 it lists for Claude Opus 5.
Where it sits
- $0.262/M$ per 1M tokensBlended price (3:1)Lower is better#19 / 87peer median $1.69/Mvendor price, checked by HokAI
- 217 tok/stokens/sOutput speedHigher is better#15 / 53peer median 94 tok/scited: Artificial Analysis
Cheaper than 77% of the 87 GA models with a published price, rank 15 of 53 on output speed as cited from Artificial Analysis, and one of 33 that document a zero-data-retention option.
Ranks are against GA models on HokAI that publish the same figure; ties share a rank.
Provider: DeepSeek · Family: DeepSeek V4
Context window: 1,048,576 tokens · Max output: 384,000
Input modalities: text, image, tool-calls · Output: text, tool-calls
About DeepSeek-V4.1-Flash-1017
DeepSeek-V4.1-Flash-1017 is the 17 October 2026 update to DeepSeek-V4.1-Flash, published by DeepSeek as open weights under the MIT licence. The company calls it a minor update that delivers stronger benchmark results, faster task completion and shorter outputs than the 10 September build. The API name deepseek-flash now serves it, and the retired names deepseek-v4-flash and deepseek-v4-flash-vision-exp are still accepted and routed to it. DeepSeek says it is live on its website, in its app and on the API. It sits beside DeepSeek-V4-Pro-0813, the larger model that stays on the same pricing page.
The release note does not restate the architecture, so the details here come from the repository config and the V4.1-Flash model card that DeepSeek links as its technical report. It is a 40-layer mixture-of-experts model with 384 routed experts, one shared expert and 6 routed experts per token. The V4.1-Flash card describes a causal encoder-decoder layout that activates 8B parameters per token during prefill and 16B during decode, with a 552B backbone and 196B parameters of Engram memory. Hugging Face counts 763B stored parameters in the 1017 repository, which also holds the vision encoder. The same card reports a global KV cache of 890 bytes per token, for a context of up to one million tokens (1,048,576 positions in config.json).
DeepSeek's table compares the model with K3, GLM-5.3, GPT-5.6 Sol (from OpenAI) and Claude Opus 5 (from Anthropic), and every figure below is vendor-reported. On repo-level coding it leads: DeepSWE v1.1 reaches 77.0 with the Mini-SWE-V2 scaffold, against 74.2 for the September build, 74.0 for Opus 5, 73.0 for GPT-5.6 Sol and 66.9 for GLM-5.3. Automation-Bench v1.0.6 (60.5 against 50.3 for Opus 5), Apex Bench v1.1 (72.1 against 65.8) and Agents' Last Exam (35.4 against 28.6) also go to this model. It trails on the hardest rows: Terminal-Bench 4.0 with Claude Code as the scaffold is 42.3 against 51.8 for Opus 5, ProgramBench is 29.1 against 37.0, NL2Repo-Bench is 68.2 against 75.3, and Humanity's Last Exam in its no-tools setting is 42.6 against 56.3 (44.0 on the text-only subset). Against the September build the gains are large: Terminal-Bench 4.0 rises from 31.2 and ProgramBench from 20.3. The vision rows run with tools, and Chartography at 83.0, BabyVision at 92.5 and ZeroBench-main at 51.0 each land within two points of the best rival score listed.
Artificial Analysis (AA) scored the maximum-reasoning configuration at 45 on its Intelligence Index v4.3.2 and measured 216.5 output tokens per second. Running that index took 210M output tokens, a volume AA's page describes as very verbose, and AA's cost per index task was $0.26. That sits uneasily with the shorter-outputs claim in the release note, which compares the model with its own September build rather than with other vendors, so test token use on your own prompts before counting on the saving.
The hosted API speaks both the OpenAI and the Anthropic formats at api.deepseek.com and api.deepseek.com/anthropic, takes the model name deepseek-flash, and lists JSON output, tool calls, the Responses API and beta chat-prefix completion as supported. Fill-in-the-middle completion (beta) works only in non-thinking mode, and thinking is the default. Maximum output is 384K tokens and the listed concurrency limit is 2,500. DeepSeek tuned the release for its own agent toolkit (v0.2.1-rc.1 or later), Tencent CodeBuddy and WorkBuddy, and documents setups for Claude Code, OpenCode and OpenClaw. OpenRouter lists the model as deepseek/deepseek-v4.1-flash-1017.
The weights ship as safetensors shards under the MIT licence, with FP8 tensors and FP4 experts per config.json. DeepSeek's recommended sampling is temperature 1.0, top_p 1.0 and max_tokens of at least 256K. The model card and release note name no system card, training-data cutoff or SLA, and the pages read state no retention period or zero-retention option for the hosted API, so regulated workloads should plan around self-hosting or a written agreement. The DeepSeek provider page lists the rest of the family, and the open-source model facet shows other open releases.
Pricing
Search and Fetch bill separately at $0.75 per 1,000 calls each, and calls made through server-side tools in the DeepSeek agent toolkit, Claude Code and Codex are free until billing starts on 31 October 2026 at 16:00 UTC. Requests to the retired legacy model names bill at the Flash rates. A cache hit costs 98% less than a miss, which suits agent runs built around one long fixed prefix. [DeepSeek-V4-Pro-0813](/hub/models/deepseek-v4-pro-0813) and [a ranking of low-cost LLM APIs](/guides/cheapest-llm-api-2026) give the nearest price context.
What a real job costs
| Job | Input | Output | Total |
|---|---|---|---|
| Summarise a 20-page PDF | $0.0045 | $0.0006 | $0.0051 |
| Support reply | $0.0003 | $0.0002 | $0.0005 |
| One coding agent run | $0.030 | $0.012 | $0.042 |
Budgets: 20-page PDF = 30k in / 1k out · Support reply = 2k in / 300 out · Coding agent run = 200k in / 20k out. Computed from the vendor's per-token prices at render time; cached-input discounts are not applied.
Key Features
- 1M-token context: The API pricing page lists a 1M-token context with a 384K maximum output, and config.json sets 1,048,576 positions.
- Image input: Vision is supported on the API, while the same page marks it unsupported for DeepSeek-V4-Pro-0813; config.json caps the vision encoder at 1,024 image tokens.
- Thinking and non-thinking modes: Thinking is the default mode and non-thinking is selectable, and the repository's encoding tests cover numeric reasoning effort.
- OpenAI and Anthropic API formats: The model name deepseek-flash works against api.deepseek.com and api.deepseek.com/anthropic, with JSON output, tool calls and the Responses API listed as supported.
- MIT-licensed weights: The Hugging Face release is 48 safetensors shards, and config.json stores the experts in FP4.
- Compressed KV cache: The V4.1-Flash card reports 890 bytes of global KV cache per token with FP4 main KV, roughly one quarter of DeepSeek-V4-Flash.
Pros
- Open MIT weights plus a hosted API give a self-host exit path that the proprietary models in DeepSeek's table cannot offer.
- Image input is available at the same per-token rates as text, with no separate vision price on the pricing page.
- The listed concurrency limit is 2,500, five times the 500 shown for DeepSeek-V4-Pro-0813, which helps parallel agent fleets.
- Cache-hit input is priced 98% below cache-miss input, which rewards long fixed prefixes in agent loops.
Cons
- Trails Claude Opus 5 on the hardest rows of the vendor table, including Humanity's Last Exam, terminal tasks and ProgramBench.
- AA measured 210M output tokens to complete its index, so real cost per task may exceed what the low per-token rates suggest.
- No audio or video input or output is listed anywhere in the release.
- Self-hosting means a 510 GB download with no published VRAM guidance, and the hosted API has no stated retention or zero-retention terms.
Benchmarks
- Humanity's Last Exam: 42.6% vendor-reported · 20 Oct 2026 — Expert-written questions across many fields, % correct.
- AA Intelligence Index: 45 cited: Artificial Analysis · 20 Oct 2026 — Composite of 10 evaluations run by Artificial Analysis, 0 to 100.
- Output speed: 217 tok/s cited: Artificial Analysis · 20 Oct 2026 — Median tokens written per second as measured by Artificial Analysis.
A benchmark is an exam, not the job. Scores transfer unevenly between tasks, so weigh the one closest to your workload and read every figure with its source.
Frequently Asked Questions
What does DeepSeek-V4.1-Flash-1017 actually cost?
DeepSeek's pricing page lists $0.15 per 1M input tokens on a cache miss, $0.003 on a cache hit and $0.60 per 1M output tokens off-peak. Peak hours (01:00-04:00 and 06:00-10:00 UTC, Monday to Friday, excluding Chinese public holidays) double every rate, to $0.30, $0.006 and $1.20. The page lists no free API tier. On the same page DeepSeek-V4-Pro-0813 costs $0.66 input and $1.98 output off-peak, so this model runs at about 23% of its input rate and 30% of its output rate.
Is DeepSeek-V4.1-Flash-1017 better than Claude Opus 5?
On DeepSeek's own table the answer depends on the task. It posts Automation-Bench v1.0.6 at 60.5 against 50.3 and Apex Bench at 72.1 against 65.8, but ProgramBench at 29.1 against 37.0 and NL2Repo-Bench at 68.2 against 75.3. Opus 5 also scores higher on all three vision rows run with tools, for example BabyVision at 94.1 against 92.5. These are vendor-reported figures on setups DeepSeek chose, so replicate on your own repositories before switching.
Is DeepSeek-V4.1-Flash-1017 open source, and can you run it yourself?
The weights are published on Hugging Face under the MIT licence, which permits commercial use. The download is 48 safetensors shards totalling about 510 GB, and the repository adds inference scripts but no training code or data. DeepSeek states no VRAM requirement, so size your hardware from the 510 GB footprint and test before committing.
Does DeepSeek train on your API prompts?
The pages read do not say. DeepSeek's privacy policy (updated 10 February 2026) covers its apps, websites and software, lists training and improving its models as a purpose, states that data is stored in China and describes a right to opt out of training. It excludes data handled inside downstream applications built on the open platform, and it names no fixed retention period or zero-retention option for API inputs. Get written confirmation for sensitive data, or self-host the MIT weights.
Who should use DeepSeek-V4.1-Flash-1017, and who should skip it?
It fits teams running high-volume coding or automation agents where cost per task counts, since AA measured $0.26 per task on its full index, and teams that want open weights with image input. Skip it if your work depends on the hardest reasoning, where the Humanity's Last Exam gap to Opus 5 is large, or if you need audio input. Also skip the hosted API if data-residency rules bar a China-hosted service and you cannot host the weights yourself.
Top Alternatives
- DeepSeek-V4.1-Flash: Pick 1017 over the September build for its higher DeepSWE and terminal-task scores in DeepSeek's table; the deepseek-flash name already serves it.
- DeepSeek-V4-Pro-0813: Pick 1017 if you need image input, a higher concurrency limit and lower per-token rates, all listed on DeepSeek's pricing page; V4-Pro-0813 is listed without vision.
- Claude Opus 5: Pick 1017 for open weights and repo-level coding scores; pick Opus 5 if the hardest reasoning and terminal tasks matter more, since it leads those rows in DeepSeek's table.
- GPT-5.6 Sol: Pick 1017 for coding-agent work and self-hosting; pick GPT-5.6 Sol if Humanity's Last Exam and ZeroBench-main matter most, where it scores higher in DeepSeek's table.
- GLM-5.3: Pick 1017 for coding agents, since DeepSeek's table has it ahead of GLM-5.3 on every text benchmark listed; keep GLM-5.3 where you have already validated it.