Here is an odd pair of numbers. In the week of October 4–10, 2026, DeepSeek V4.1 Flash was the most-used text model on OpenRouter, at roughly 37.5 trillion tokens. On Arena's human-preference leaderboard from the same week, its max entry ranked 39th overall.
Both are true, and the price of deepseek flash, as DeepSeek's API calls it, explains most of the gap. Since October 17 that API name runs a new build, DeepSeek Flash 1017 (officially DeepSeek-V4.1-Flash-1017), with higher benchmark scores at the same list price.
What Changed in DeepSeek V4.1-Flash 1017
DeepSeek released DeepSeek-V4.1-Flash-1017 on October 17, 2026. Its Hugging Face model card calls it "a minor update to DeepSeek-V4.1-Flash" and says it also cuts task execution time and output length. DeepSeek's announcement says the website, app and API switched to the new model the same day.
There is no new model ID. The Models & Pricing page still uses deepseek-flash and now lists DeepSeek-V4.1-Flash-1017 as its model version, at the same prices shown below. Requests to the legacy names deepseek-v4-flash and deepseek-v4-flash-vision-exp are also served by 1017 and billed at Flash rates.
DeepSeek's own benchmark table, previous V4.1 Flash against 1017:
| Benchmark | V4.1 Flash | V4.1-Flash-1017 | Change |
|---|---|---|---|
| HLE, no tools | 36.8 | 42.6 | +5.8 |
| Terminal-Bench 4.0 | 31.2 | 42.3 | +11.1 |
| DeepSWE v1.1 | 74.2 | 77.0 | +2.8 |
| ProgramBench | 20.3 | 29.1 | +8.8 |
| NL2Repo-Bench | 65.4 | 68.2 | +2.8 |
| Automation-Bench v1.0.6 | 54.8 | 60.5 | +5.7 |
| Agents' Last Exam | 31.8 | 35.4 | +3.6 |
| Apex Bench v1.1 | 64.6 | 72.1 | +7.5 |
| Chartography (with tools) | 78.9 | 83.0 | +4.1 |
| BabyVision (with tools) | 89.6 | 92.5 | +2.9 |
| ZeroBench-main (with tools) | 49.0 | 51.0 | +2.0 |
DeepSeek ran Terminal-Bench 4.0 in the Claude Code harness and DeepSWE v1.1 in the Mini-SWE-V2 harness. On HLE's text-only subset the scores are 39.1 and 44.0. The announcement names DeepSeek Harness, Tencent CodeBuddy and WorkBuddy as tuning targets and recommends DeepSeek Harness v0.2.1-rc.1 or later.
The weights are open under the MIT license on Hugging Face and ModelScope. Hugging Face lists 763B parameters. The release has no Jinja chat template; prompt encoding comes as a Python reference implementation in the encoding folder, plus the deepseek-recipe Rust libraries. Recommended sampling is temperature 1.0 and top_p 1.0, with max_tokens of at least 256K.
The same release opened DeepSeek's Search and Fetch APIs to everyone. Each is billed separately at $0.75 per 1,000 successful requests. If you use the DeepSeek API through DeepSeek Harness, Claude Code or Codex, server-side tool calls already hit these endpoints. Those calls are free for now; DeepSeek will start billing them on October 31, 2026, at 16:00 UTC.
DeepInfer Runs V4.1 Flash Offline
On October 16, 2026, DeepSeek open-sourced DeepInfer, the inference framework behind its production API, under Apache 2.0. The release supports a single model, DeepSeek-V4.1-Flash, and by default it downloads the DeepSeek-V4.1-Flash-1017 checkpoint from Hugging Face on first run. The repository had 19,190 stars and 2,017 forks on October 22.
It targets data-center hardware only: NVIDIA SM100 (Blackwell) GPUs and Huawei Ascend 950 NPUs. The build needs Linux, GCC 13 or newer, Python 3.12, and CUDA 13 or CANN 9.2. DeepSeek's published throughput curves were measured on 4 SM100 GPUs for prefill and 64 for decode.
| Feature in the release | What it does |
|---|---|
| FlashMLA, DeepGEMM, DeepEP, TileLang kernels | DeepSeek's own attention, matmul and expert-parallel kernels |
| Attention DP, MoE EP + EPLB | Data parallelism for attention, expert parallelism with load balancing |
| DSpark speculative decoding | Draft tokens with confidence-scheduled verification, up to 5 by default |
| SWA bounded replay | Replays only a short suffix during prefill (140 tokens by default for V4.1) |
| Batch-invariant inference, deterministic sampling | Identical outputs regardless of batch mix (NVIDIA only) |
| Structured outputs | Regex, JSON schema or EBNF constraints |
The open release covers offline inference. DeepSeek lists what it leaves out: a KV cache system, prefill/decode disaggregation, and load balancing across data-parallel replicas. Teams that want an online service on their own hardware still have to build those parts or use another serving stack. The README credits vLLM as DeepInfer's original base, with ideas from SGLang and TensorRT-LLM.
→ Track it on AI Rank: Open-source repository ranking
DeepSeek V4.1 Flash Pricing
DeepSeek released V4.1 Flash on September 10, 2026. On the official API the model name is deepseek-flash. List prices from the Models & Pricing page, unchanged between October 12 and October 17, 2026, in USD per million tokens:
| Off-peak | Peak | |
|---|---|---|
| Input, cache hit | $0.003 | $0.006 |
| Input, cache miss | $0.15 | $0.30 |
| Output | $0.60 | $1.20 |
Peak means weekdays 01:00–04:00 and 06:00–10:00 UTC, excluding Chinese public holidays. Everything else, weekends included, is half price. A cache hit costs one-fiftieth of a miss, so repetitive prompts get very cheap. That's the short answer for anyone who searched deepseek v4.1 flash pricing.
You don't give up much for it: a 1M-token context, up to 384K output tokens, native vision, thinking and non-thinking modes (thinking by default) and a concurrency limit of 2,500. On the same page, deepseek-v4-pro (V4-Pro-0813) costs $0.66 / $1.98 off-peak and $1.32 / $3.96 at peak, caps concurrency at 500 and has no vision.
Legacy IDs deepseek-v4-flash and deepseek-v4-flash-vision-exp still work, but the original V4 Flash models are retired; their requests now go to DeepSeek-V4.1-Flash-1017 at Flash rates.
Usage Ranking
| Window | Rank by tokens | Tokens | Change |
|---|---|---|---|
| Day, Oct 10, 2026 | #1 | ~4.84T | −12% |
| Week, Oct 4–10 | #1 | ~37.5T | +54% |
| Month, Sep 11–Oct 10 | #1 | ~95.2T | >999% |
| Week, Oct 10–16 | #1 | ~34.8T | −4% |
That's deepseek/deepseek-v4.1-flash-20260910 on OpenRouter's public ranking (CC BY 4.0 via AI Rank). OpenRouter labels it "(batch)" even though it isn't the :batch route; the real :batch ID is a separate, much smaller row. These figures cover the September 10 build; the newest window, Oct 10–16, ends the day before 1017 went live.
Arena Preference Scores
Arena's text leaderboard, style control, snapshot of October 8, 2026, entry deepseek-v4.1-flash-max (license listed as MIT):
| Category | Rank | Rating (95% CI) | Votes |
|---|---|---|---|
| Overall | #39 | 1474.5 (1468.0–1481.1) | 10,872 |
| Coding | #26 | 1527.6 (1516.6–1538.6) | 3,026 |
| Instruction following | #23 | 1476.5 (1466.5–1486.4) | 3,826 |
| Math | #23 | 1493.3 (1469.9–1516.7) | 579 |
| Chinese | #31 | 1518.3 (1497.9–1538.8) | 856 |
| Creative writing | #59 | 1439.3 (1425.9–1452.6) | 2,224 |
| Multi-turn | #69 | 1465.6 (1450.5–1480.7) | 1,581 |
Solid on coding, instructions and math; weaker on creative writing and long conversations. Math rests on under 600 votes, so take it loosely. This snapshot also predates the 1017 update.
→ Track it on AI Rank: DeepSeek V4.1 Flash max — Arena rating, rank and 90-day history
Put together, the picture is a workhorse: the model people send bulk jobs and agent loops to when cost per token matters more than polish. Live positions are on the OpenRouter model usage ranking on AI Rank.
Compare before you choose: pricing changes and rankings move daily — check the current model rankings on AI Rank before you commit.
Or 📱 get the AI Rank app to follow them on your phone.
Sources
- DeepSeek: DeepSeek-V4.1-Flash-1017 announcement
- DeepSeek-V4.1-Flash-1017 on Hugging Face
- DeepSeek-V4.1-Flash-1017 on ModelScope
- DeepSeek Models & Pricing
Prepared with automated editorial assistance and checked against official sources and dated AI Rank data on October 12, 2026; the 1017 section was added on October 17, 2026 from DeepSeek's announcement, model card and pricing page; the DeepInfer section and the Oct 10–16 usage row were added on October 22, 2026 from the DeepInfer repository and AI Rank data. We did not run our own benchmarks for this article.