The Flash checkpoint released on 17 October 2026. Everything below comes from three primary sources — the official model card, the official API price list and DeepSeek's own announcement: what changed, the benchmark numbers, the per-token price, and how to get the weights. Anything we could not verify is labelled as such.
DeepSeek-V4.1-Flash-1017 is the Flash-tier checkpoint released on 17 October 2026. DeepSeek calls it a minor update to V4.1-Flash, but most of the agentic and coding rows on the model card move by double digits (Terminal-Bench 4.0 from 31.2 to 42.3, ProgramBench from 20.3 to 29.1), while the vendor also reports shorter task times and shorter outputs. In other words: stronger and cheaper to run, not just a higher score. The weights stay MIT-licensed, the API model name is still deepseek-flash, and the price list did not change.
Both columns come from the same official model card: the new checkpoint on the left, the September 2026 V4.1-Flash on the right. Same card, same harnesses, so the two columns can be subtracted directly.
| Benchmark | V4.1-Flash-1017 | V4.1-Flash (Sep 2026) | Change |
|---|---|---|---|
| Terminal-Bench 4.0 | 42.3 | 31.2 | +11.1 |
| ProgramBench | 29.1 | 20.3 | +8.8 |
| Apex Bench v1.1 | 72.1 | 64.6 | +7.5 |
| HLE without tools | 42.6 | 36.8 | +5.8 |
| Automation-Bench v1.0.6 | 60.5 | 54.8 | +5.7 |
| HLE (text-only subset) | 44.0 | 39.1 | +4.9 |
| Chartography (with tools) | 83.0 | 78.9 | +4.1 |
| Agents' Last Exam | 35.4 | 31.8 | +3.6 |
| DeepSWE v1.1 | 77.0 | 74.2 | +2.8 |
| NL2Repo-Bench | 68.2 | 65.4 | +2.8 |
| BabyVision (with tools) | 92.5 | 89.6 | +2.9 |
| ZeroBench-main (with tools) | 51.0 | 49.0 | +2.0 |
Every number below is copied from the single comparison table on the official model card — nothing rescaled, nothing interpolated. Bold marks the highest score among the five models in that row; cyan marks the model this page is about.
| Benchmark | V4.1-Flash-1017 | V4.1-Flash | K3 | GLM-5.3 | GPT-5.6 Sol | Opus 5 |
|---|---|---|---|---|---|---|
| Terminal-Bench 4.0 | 42.3 | 31.2 | 12.6 | 37.9 | 39.9 | 51.8 |
| DeepSWE v1.1 | 77.0 | 74.2 | 67.5 | 66.9 | 73.0 | 74.0 |
| ProgramBench | 29.1 | 20.3 | 17.5 | 19.0 | 23.0 | 37.0 |
| NL2Repo-Bench | 68.2 | 65.4 | 58.0 | 58.0 | 56.8 | 75.3 |
| Automation-Bench v1.0.6 | 60.5 | 54.8 | 46.7 | 48.8 | 45.8 | 50.3 |
| Agents' Last Exam | 35.4 | 31.8 | 27.6 | 28.5 | 26.7 | 28.6 |
| Apex Bench v1.1 | 72.1 | 64.6 | 47.8 | 56.6 | 54.3 | 65.8 |
| HLE (no tools) | 42.6 (44.0*) | 36.8 (39.1*) | 43.5 | 42.0* | 44.5 | 56.3 |
| Chartography (with tools) | 83.0 | 78.9 | 68.1 | — | 79.9 | 84.0 |
| BabyVision (with tools) | 92.5 | 89.6 | 85.7 | — | 88.9 | 94.1 |
| ZeroBench-main (with tools) | 51.0 | 49.0 | 41.0 | — | 53.0 | 52.0 |
The 1017 checkpoint is served under the same API model name, deepseek-flash, at the same list price as the previous Flash version. The table below is taken from the official price list, per 1M tokens.
| Line item (per 1M tokens) | Off-peak | Peak |
|---|---|---|
| Input · cache hit | $0.003 | $0.006 |
| Input · cache miss | $0.15 | $0.30 |
| Output | $0.60 | $1.20 |
| Time-of-day rule | Peak = Mon-Fri 01:00-04:00 and 06:00-10:00 UTC (excluding Chinese public holidays); everything else, including weekends, is half-price off-peak. | |
Same checkpoint, different routes, different prices. The table below is what OpenRouter listed on 2026-10-17 — third-party marketplace data that can change at any time.
| Route | Input / 1M | Output / 1M | Cache read / 1M |
|---|---|---|---|
| inference.net | $0.10 | $0.50 | $0.002 |
| Cloudflare | $0.10 | $0.60 | $0.003 |
| DeepSeek（官方） | $0.15 | $0.60 | $0.003 |
| DeepInfra（-30%） | $0.14 | $0.42 | $0.0042 |
The weights are still MIT-licensed and the repository carries inference and encoding instructions, but there is one trap this time: no Jinja chat template.
Published on Hugging Face (deepseek-ai/DeepSeek-V4.1-Flash-1017) and ModelScope under the MIT licence, tagged FP8 8-bit and image-text-to-text. Within hours of release the HF repo had 680 likes and 11 open discussions.
This release ships no Jinja-format chat template. The repository's encoding folder contains a self-contained Python reference implementation covering multi-turn conversations, tool calling, thinking mode, numeric reasoning effort, mid-conversation system messages and interleaved image content.
DeepSeek also publishes deepseek-recipe, a set of Rust libraries with Python bindings: it converts Messages, Chat Completions and Responses API requests into the Conversation format, encodes them into prompts or token IDs, and parses output back into complete or streamed responses. Model inference, tool execution and HTTP transport are left to the caller.
The model card's recommended defaults: temperature 1.0, top_p 1.0, a 1M-token context window, and max_tokens of at least 256K. That is a minimum output budget, not a target length to fill.
It is the release date encoded in the model name: 17 October 2026. DeepSeek describes it as a minor update to DeepSeek-V4.1-Flash rather than a new generation, but the model card shows double-digit gains on several agentic benchmarks, for example Terminal-Bench 4.0 rising from 31.2 to 42.3.
On DeepSeek's own API it is deepseek-flash; there is no dated model ID to call. The legacy names deepseek-v4-flash and deepseek-v4-flash-vision-exp are still accepted, but those models have been retired: the requests are served by DeepSeek-V4.1-Flash-1017 and billed at the Flash price. OpenRouter lists the same checkpoint as deepseek/deepseek-v4.1-flash-1017.
Official DeepSeek list prices per 1M tokens: cache-hit input $0.003 off-peak and $0.006 at peak, cache-miss input $0.15 off-peak and $0.30 at peak, output $0.60 off-peak and $1.20 at peak. Off-peak is half of peak; peak hours are 01:00-04:00 and 06:00-10:00 UTC, Monday to Friday, excluding Chinese public holidays. Third-party routes on OpenRouter start at $0.10 input and $0.50 output per 1M tokens.
Yes. The weights are published under the MIT licence on Hugging Face and ModelScope. Note that this release ships no Jinja-format chat template: DeepSeek provides a Python reference encoder in the repository encoding folder and a Rust toolkit called deepseek-recipe for production prompt formatting. Recommended sampling settings are temperature 1.0, top_p 1.0 and at least 256K max output tokens.
Yes to both. The checkpoint is tagged image-text-to-text, and DeepSeek's pricing table lists vision support on Flash, while V4-Pro does not take images. Tool calls, JSON output, the Responses API and the Anthropic API format are all listed as supported, with a concurrency limit of 2,500.
Page data was read from the primary sources below on 2026-10-17. Prices, benchmark numbers and limits change — check the vendor pages before you commit.