Primary use stated or inferred cautiously from official documentation.
D
DEEPSEEK / deepseek-flash
DeepSeek V4.1 Flash
DeepSeek updated its API, website and app to DeepSeek-V4.1-Flash-1017 on October 17, 2026. The API identifier remains deepseek-flash and pricing is unchanged. We retain the profile and its URL: this is a revision of the existing model, not another duplicate model.
01 / SUMMARY
The essential
Maximum input capacity when published by the source.
Documented maximum generation limit.
Verified availability channels.
Primary functional family and verified specialties.
Declared terms for API access, model weights, or self-hosting.
02 / TECHNICAL SHEET
Limits and integration
| API ID | deepseek-flash |
|---|---|
| Model type | Text and reasoning |
| Access model | Paid API · open weights |
| License | MIT |
| Deployment | Hosted API · self-hosted infrastructure |
| Release | 10 SEP 2026 |
| Knowledge cutoff | Not published |
| Entrance | Text · Image |
| Exit | Text |
| Context window | 1,000,000 tokens |
| maximum output | 384.000 tokens |
| Reasoning | Modes with and without reasoning |
| Published tools | Features · Responses API · Chat prefix · FIM |
| Structured Outings | Supported |
| Batch processing | Not published |
| Prompt cache | Supported |
| Fine-tuning | Open weights |
| Verified platforms | DeepSeek API · Hugging Face |
Agents, coding, and open deployment at low cost.
DeepSeek-V4.1-Flash-1017 weights are available in the official repository under the MIT license. The October 17 results appear in both the API update and the checkpoint card: these are vendor measurements, not tests run by Inferama. GPQA and Terminal-Bench 2.1 retain their historical September 10 scores. The local release does not include a Jinja chat template: consult the official encoder and review its code before execution. Downloadable weights do not mean free inference.
03 / CAPABILITIES
What it can do
Orientation
Agents, coding, and open deployment at low cost.
Context
1.000.000 tokens · 384.000 tokens
Tools and integration
Features · Responses API · Chat prefix · FIM
Access
DeepSeek API · open weights
Input modalities
TextImage
04 / PRICES
Documented cost
| Concept | Worth | Unit/condition |
|---|---|---|
| Standard input | 0,30 USD / 1 M tokens (peak) | Standard API rate |
| Cached input | 0,006 USD / 1 M tokens (peak) | Reading reused prefixes |
| Cache write or storage | Not published | The condition varies by provider |
| Standard output | 1,20 USD / 1 M tokens (peak) | May include reasoning tokens |
| Batch input | Not published | Asynchronous processing |
| Batch output | Not published | Asynchronous processing |
Off-peak rates are half price; schedules are published in UTC.
i Prices change and may depend on level, region or context length. Check the source before making a decision.
05 / EVALUATIONS
How to read the results
“
DeepSeek publishes twelve new results, reporting full HLE and its text-only subset separately. Terminal-Bench 4.0 uses Claude Code; DeepSWE v1.1 uses Mini-SWE-V2. Chartography, BabyVision and ZeroBench use tools; ZeroBench-main measures Pass@5. These are not evaluations run by Inferama or direct comparisons between different protocols.
| Benchmark | Result | Metric | Source |
|---|---|---|---|
| GPQA Diamond | 90,9 % | Accuracy | View source ↗ |
| Terminal-Bench 2.1 | 90,6 % | Accuracy | View source ↗ |
| HLE (sin herramientas) | 42,6 | Score | View source ↗ |
| HLE (texto, sin herramientas) | 44,0 | Score | View source ↗ |
| Terminal-Bench 4.0 (Claude Code) | 42,3 | Score | View source ↗ |
| DeepSWE v1.1 | 77,0 | Score | View source ↗ |
| ProgramBench | 29,1 | Score | View source ↗ |
| NL2Repo-Bench | 68,2 | Score | View source ↗ |
| Automation-Bench v1.0.6 | 60,5 | Score | View source ↗ |
| Agents’ Last Exam | 35,4 | Score | View source ↗ |
| Apex Bench v1.1 | 72,1 | Score | View source ↗ |
| Chartography (con herramientas) | 83,0 | Score | View source ↗ |
| BabyVision (con herramientas) | 92,5 | Score | View source ↗ |
| ZeroBench-main (Pass@5, herramientas) | 51,0 | Score | View source ↗ |
Inferama only highlights a “best result” when the metric, test set, configuration, and date allow for an equivalent comparison. The supplier's figures are presented as claims from its own source.
06 / VERSIONS
Chronology
DeepSeek-V4.1-Flash-1017
API, website, app and public checkpoint under MIT. Weight identifier: deepseek-ai/DeepSeek-V4.1-Flash-1017. The deepseek-flash alias and pricing remain unchanged.
DeepSeek V4.1 Flash
Original multimodal release. DeepSWE v1.1: 74.2; GPQA Diamond: 90.9; Terminal-Bench 2.1: 90.6. Do not mix with revision 1017.
ANALYSIS
Related analysis
DeepSeek V4.1 Flash: How to Test Whether Its Compressed Cache Reduces an Agent’s Real Cost
DeepSeek says V4.1 Flash has a much smaller persistent KV cache than its predecessor. That figure describes an infrastructure property, not the cost or completion time of an entire task. This analysis proposes a controlled test to measure both.
28 Sep 2026 ↗
DeepSeek V4.1 Flash + Eleven v3 vs. full-duplex voice: how to compare the architectures
A modular chain built from speech recognition, DeepSeek V4.1 Flash, and Eleven v3 is not equivalent to a full-duplex voice model. This guide proposes comparing both systems end to end, without declaring a winner based on isolated specifications.
26 Sep 2026 ↗
DeepSWE v1.1: What a Patch That Passes Its Verifier Measures—and When Its pass@1 Stops Showing That the Agent Solved the Task
DeepSWE v1.1 evaluates code changes produced by agents on original, long-horizon tasks. Its result can provide a useful signal, but a leaderboard row is interpretable only when the harness, error policy, rollouts, verifier, and exact benchmark version are known. A review by Epoch AI adds a specific warning: under certain evaluation conditions, some patches that modify tests can receive a false negative.
23 Sep 2026 ↗07 / SOURCES
Traceability
LAUNCHDeepSeek · actualización del 17 de octubreapi-docs.deepseek.com↗
APIDeepSeek · versión, límites y preciosapi-docs.deepseek.com↗
PESOSDeepSeek V4.1 Flash · pesos del lanzamiento originalhuggingface.co↗
PESOSDeepSeek · checkpoint 1017 and MIT licensehuggingface.co↗
LICENSEMIT · DeepSeek-V4.1-Flash-1017huggingface.co↗