DeepSeek V4.1 Flash: Specs, Pricing, Context Window and Availability Facts

DeepSeek V4.1 Flash is an open-weight, multimodal mixture-of-experts language model from DeepSeek, released on 10 September 2026 and updated to DeepSeek-V4.1-Flash-1017 on 17 October 2026, served through the API name deepseek-flash. This profile collects the published facts about it in one place: API name, context window, output limit, architecture, peak and off-peak prices, features, availability, the Hugging Face model cards, data terms and the retirement of the V4 Flash names it replaced. It is written for developers, architects, buyers and governance teams who need checked numbers rather than launch coverage. Every fact comes from DeepSeek's API documentation, change log, Hugging Face model cards and policies, listed under Standards and references. Where DeepSeek publishes nothing on a point, the page says so.

Last checked: 2026-11-15.

This is an independent reference page from NHI Mgmt Group based on public provider documentation; it is not sponsored or endorsed by any provider. It uses DeepSeek's public API documentation as of November 2026. Model details change, so confirm against DeepSeek's current documentation.

Key takeaways

• One API name, latest model. DeepSeek's change log tells callers to "simply use deepseek-flash to access the latest model"; since 17 October 2026 that is DeepSeek-V4.1-Flash-1017.
• 1M tokens in, up to 384K out. The pricing page lists a 1M-token context length, output up to 384K, thinking mode on by default and vision input.
• Prices depend on time of day. Cache-miss input costs $0.30 at peak and $0.15 off-peak per million tokens, output $1.20 and $0.60, with peak hours set in UTC on weekdays.
• Open weights under MIT. The Hugging Face cards state the MIT License for the repository and weights, with 763B total parameters in the Safetensors metadata and about 8B to 16B active per token.
• It replaced V4 Flash. V4 Flash and V4 Flash Vision Exp were retired on 10 September 2026, and their API names are temporarily routed to V4.1 Flash at the Flash price.

What is DeepSeek V4.1 Flash?

DeepSeek's change log entry for 10 September 2026 describes DeepSeek-V4.1-Flash as the smallest model in DeepSeek's new architecture, with native multimodal visual understanding, and on the same date moved callers to the deepseek-flash model name and reduced API prices. The Hugging Face card describes a multimodal mixture-of-experts model that takes image and text input and returns text, with a vision encoder, DeepSeek-ViT, trained from scratch.

The 17 October 2026 update, DeepSeek-V4.1-Flash-1017, is described on its card as a "minor update" with improvements in reasoning, coding, agentic tasks and visual benchmarks, aiming to reduce task execution time and output length. The change log states that the calling method and pricing did not change. DeepSeek publishes the weights, so the model can be called through the DeepSeek API or self-hosted. DeepSeek does not state a knowledge cutoff on the pages checked. For the full DeepSeek line-up, see DeepSeek Models.

Key facts at a glance

As of 2026-11-15. Source: DeepSeek pricing page, change log and Hugging Face model cards.

Release and version history

As of 2026-11-15. Dates from the DeepSeek API change log.

The four-digit suffix in DeepSeek-V4.1-Flash-1017 is the month and day of the update. Because deepseek-flash always points to the latest Flash version, an application calling that name moved to the 1017 weights on 17 October without a code change. To fix behaviour to one set of weights, self-host the published checkpoint.

How much does DeepSeek V4.1 Flash cost?

As of 2026-11-15. Price per million tokens on the DeepSeek API. The pricing page shows amounts with a dollar sign and does not name the currency. Source: DeepSeek models and pricing page.

Peak and off-peak pricing took effect at 16:00 UTC on 16 August 2026, according to the change log. DeepSeek notes that it may adjust prices and recommends checking the page regularly. Worked example at peak rates with no cache hits: a request with 100,000 input tokens and 10,000 output tokens costs $0.03 for input plus $0.012 for output, $0.042 in total; the same request off-peak costs $0.021.

Context window and output limits

The context length is 1M tokens and output can run to 384K tokens. The 1017 card recommends setting max_tokens to at least 256K, with temperature 1.0 and top_p 1.0, so long reasoning traces are not cut short. Plan cost on output as well as input, since thinking mode is the default and output is priced at four times cache-miss input.

Modalities and tools

V4.1 Flash takes text and image input and returns text. The V4.1-Flash card describes a 40-layer Transformer split into a 20-layer causal encoder and a 20-layer decoder, Compressed Sparse Attention 2, and a mixture-of-experts design with 1 shared and 384 routed experts per layer, 6 routed experts active per token, plus a 196B-parameter Engram conditional memory and DSpark speculative decoding. See Mixture of Experts Explained.

• API features: JSON output, tool calls, the Responses API, the Anthropic API format and chat prefix completion (beta); FIM completion (beta) in non-thinking mode only.
• Application APIs: from 13 November 2026, web search, file input, URL reading, model safety, ASR and TTS can be enabled for deepseek-flash, with search and fetch, OCR, ASR and TTS also offered as standalone APIs.
• Chat format for self-hosting: the 1017 card provides no Jinja chat template; it ships a Python reference encoder with test cases for tool calling, thinking mode, numeric reasoning effort and interleaved images.
• Serving: the card lists vLLM, SGLang, Transformers and Docker Model Runner, and quantised builds for llama.cpp, Ollama and LM Studio.

Tool calls and URL reading let the model act on external content. Treat each tool as an integration with its own permissions and keep API keys out of code. See the LLMjacking Guide.

Where is DeepSeek V4.1 Flash available?

As of 2026-11-15. Source: DeepSeek documentation, Hugging Face and Mistral documentation.

Some governments restrict DeepSeek services on official systems. See the DeepSeek Federal Exclusion Orders and FASCSA Guide.

Access restrictions

V4.1 Flash is not a restricted-access model. The API name is listed on the public pricing page with a published concurrency limit, and the weights are downloadable from Hugging Face under MIT with no access form stated on the card. The pages checked state no tiered access programme or usage application for this model.

Benchmarks (provider-reported, dated)

As of 2026-11-15. All figures are provider-reported by DeepSeek on its Hugging Face model cards (V4.1-Flash card and V4.1-Flash-1017 card). Comparator models are those DeepSeek chose. Independent leaderboard results are not included because none were checked against their own published pages for this profile.

The 1017 card also reports Kimi K3 and GLM-5.3 as comparators. Higher is better for every score shown. Harness choices differ by benchmark, so compare only like-for-like rows.

Model card and system card

DeepSeek publishes Hugging Face model cards for both V4.1-Flash and the 1017 update, and a technical report (DeepSeek_V41_Tech_Report.pdf) in the V4.1-Flash repository. DeepSeek does not publish a separate system card or safety report for this model on the pages checked. What the cards state:

• Training: 45T tokens of multimodal pre-training, with context extended to 1M tokens at 34T tokens.
• Architecture: causal encoder-decoder, CSA2 attention with a hierarchical sparse indexer, FP4 KV caching and 384 routed experts per layer.
• Parameters: 552B backbone in card text and 763B total in Safetensors metadata, which includes components such as the Engram memory.
• Weights: Safetensors in BF16, F32, F8_E4M3 and I8 tensor types (1017 card).
• Licence: MIT for repository and weights.

Without a provider safety report, organisations that self-host need their own evaluation. See the Open-Weight Model Security Guide.

Data handling

DeepSeek's terms and privacy policy apply to API use; neither has a V4.1 Flash specific section.

• Location: the privacy policy (last updated 10 February 2026) states that DeepSeek collects, processes and stores personal data in the People's Republic of China.
• Training use: the privacy policy states inputs may be used to train and improve DeepSeek's models, and lists a right to opt out, exercised by request to DeepSeek rather than a described setting. The open platform terms do not address training on API inputs.
• Retention: the privacy policy keeps account data and inputs for as long as the account exists; the open platform terms do not address retention.
• Ownership and law: the open platform terms (effective 29 April 2026) assign output rights to the customer, let customers use inputs and outputs to train other models, and are governed by the laws of mainland China.
• Self-hosting: running the MIT weights on your own infrastructure keeps prompts within your environment. See the Self-Hosted LLM Security Guide.

For vendor risk review, see the DeepSeek and High-Risk AI Vendor Assessment Guide.

Deprecation status

As of 15 November 2026, DeepSeek lists no deprecation for deepseek-flash. The original V4.1-Flash weights were superseded on the API by the 1017 update, and the V4 Flash names it replaced are routed to it. DeepSeek describes the routing as temporary, so watch the change log.

How DeepSeek V4.1 Flash differs from other models in its family

As of 2026-11-15. Source: DeepSeek pricing page, change log and Hugging Face cards. Models listed by release date.

The documented differences from V4-Pro are price, image input, the newer V4.1 architecture and the concurrency limit (2,500 for deepseek-flash, 500 for deepseek-v4-pro). See DeepSeek V4 (Pro and Flash) and DeepSeek vs Qwen vs Kimi vs GLM.

Practitioner checklist

• Decide whether a floating name suits you: deepseek-flash changes weights without notice in code, so record the model version string returned and retest after each change log entry.
• Replace deepseek-v4-flash and deepseek-v4-flash-vision-exp in code now, since DeepSeek calls the routing temporary.
• Schedule batch-style workloads off-peak, where every rate is half the peak rate.
• Set max_tokens to at least 256K for thinking workloads, as the card recommends, and budget output cost accordingly.
• Run a data residency and vendor risk review before sending personal or regulated data to the hosted API; consider self-hosting the MIT weights instead.
• When self-hosting, verify checkpoints against the official deepseek-ai repository and use the provided reference encoder rather than an improvised chat template.
• Store DeepSeek API keys in a secrets manager and set spend alerts on the account balance.
• Check the change log monthly for updates, price changes and retirements.

Standards and references

• DeepSeek: Models and pricing (accessed 2026-11-15)
• DeepSeek: API change log (accessed 2026-11-15)
• DeepSeek on Hugging Face: DeepSeek-V4.1-Flash-1017 model card (accessed 2026-11-15)
• DeepSeek on Hugging Face: DeepSeek-V4.1-Flash model card (accessed 2026-11-15)
• DeepSeek: Open Platform Terms of Service (accessed 2026-11-15)
• DeepSeek: Privacy Policy (accessed 2026-11-15)
• Mistral AI: Models overview (accessed 2026-11-15)

Related NHI Mgmt Group resources: DeepSeek Models, DeepSeek V4 (Pro and Flash), Frontier AI Models Compared, Open-Weight Model Security Guide, How to set spend limits on LLM API keys.