Jev AI · Model research
DeepSeek V4.1
Flash 1017
What changed. What it costs.
What the benchmarks actually show.
ReDiagram Playground · Coming soon
Independent research. Not affiliated with DeepSeek. Verified 2026-10-18.
42.3
Terminal-Bench 4.0
31.2 previously
$0.15
Input / 1M tokens
Off-peak · cache miss
1M
Context window
Official specification
A focused update to V4.1 Flash.
The October 17 release updates DeepSeek’s text-and-vision model for reasoning, coding and agent tasks. DeepSeek reports better benchmark results alongside shorter task runtimes and outputs. The API calling method and pricing remain unchanged.
The official API currently maps deepseek-flash to 1017. It is a moving alias, not a verified snapshot-pinned endpoint. Open weights are published under MIT; hosted API use is billed separately.
Where the scores improved.
Official BenchmarkPublisher-reported results against the previous V4.1 Flash release. The change column is an arithmetic score difference, not a relative performance percentage.
| Benchmark | V4.1 Flash | 1017 | Change |
|---|---|---|---|
| HLE (without 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 |
| Agents’ Last Exam | 31.8 | 35.4 | +3.6 |
Read the testing conditions
Terminal-Bench 4.0 uses Claude Code; DeepSWE v1.1 uses Mini-SWE-V2. HLE above is the full reported no-tools score. The text-only subset is 44.0 versus 39.1.
What this does not establish
Different rows measure different tasks. A higher score does not guarantee better results on your workload. TryJevAI has not run an independent benchmark suite.
Independently Tested: no results available.
The same model. Two price windows.
Official DeepSeek API prices in USD per million tokens. The 1017 update did not change pricing.
| Token type | Off-peak | Peak |
|---|---|---|
| Input · cache hit | $0.003 | $0.006 |
| Input · cache miss | $0.15 | $0.30 |
| Output | $0.60 | $1.20 |
A small-request estimate
For 1,000 uncached input tokens + 500 output tokens: $0.00045 off-peak or $0.00090 peak. Calculated from the rates above; this is not a measured run or an invoice. Actual cost depends on billed usage, caching and timing.
Peak hours: weekdays 01:00–04:00 and 06:00–10:00 UTC, excluding Chinese public holidays. Other times are off-peak. Confirm current rates with DeepSeek before use.
Speed: the evidence and the gap.
What DeepSeek reports
The release notes describe lower task runtime and shorter output than the previous version. They do not provide a universal latency or tokens-per-second guarantee that applies to every API request.
What we have measured
No independent speed measurements yet. We do not assign a speed score or turn publisher claims into measured milliseconds. Future ReDiagram results should distinguish time to first token, total elapsed time and actual token usage.
API essentials, without the guesswork.
- Base URL
- https://api.deepseek.com
- Model ID
- deepseek-flash
- Version caveat
- The alias currently serves 1017 and may change later.
Use your own official provider key on your server. Open weights require a compatible inference environment; minimum self-hosting hardware for this checkpoint has not been independently verified here.
OpenCode integration
OpenCode documents a DeepSeek provider through /connect and model selection through /models. Check the actual API ID rather than assuming a “1017” display name pins a snapshot. Follow the specific setup guide for your installed version.
A few things worth knowing.
What is DeepSeek V4.1 Flash 1017?
It is the October 17, 2026 update to DeepSeek V4.1 Flash. DeepSeek documents improvements in reasoning, coding, agent tasks and visual understanding. The official API currently serves it through deepseek-flash, a moving alias rather than a guaranteed pinned snapshot.
How do I use its API?
Create your own key in the DeepSeek console. Send a server-side Chat Completions request to https://api.deepseek.com/chat/completions with model deepseek-flash. See the official API quickstart for request examples. Never put a provider key in browser code. Jev provides research; online testing is planned on ReDiagram.
Is it open source?
DeepSeek publishes the model weights and repository under the MIT license. See the official model card and license. Hosted API usage is billed separately; open weights do not mean free hosted inference.
How do I use it with OpenCode?
Use /connect to add your DeepSeek key, then /models to choose the available Flash model. If your catalog has not caught up, follow the linked official custom-provider tutorial with the actual API ID deepseek-flash. The 1017 snapshot name is not a separately verified API ID.
How does it differ from the previous version?
The official release reports higher benchmark scores and reduced task runtime and output length versus V4.1 Flash. API calling conventions and pricing did not change for the 1017 update. These are publisher reports, not independent tests by TryJevAI.
What are its API pricing and deployment requirements?
Official prices are listed on this page in USD per million tokens, with peak and off-peak rates. Using the hosted API requires a DeepSeek account, key and available balance, but no local model hardware. Self-hosting requires the published weights and a compatible inference environment; we have not verified a minimum GPU or memory configuration for this checkpoint.
Research sources: release notes, model card, pricing and OpenCode documentation. Checked 2026-10-18.
Research here. Try it on ReDiagram.
ReDiagram Playground · Coming soon
Jev provides independent model information. The online Playground will be hosted on ReDiagram. Paid API access through our sites is not available yet.