Join our Newsletter — 33% off our NHI Course
Home› Guides› DeepSeek V4.1 vs V4 vs R1: What Changed,…
Guide AI Platforms

DeepSeek V4.1 vs V4 vs R1: What Changed, Specs and Pricing Compared

PDF for newsletter subscribers · free to read online

← All guides
By NHI Mgmt Group, edited by Lalit Choda Updated 15 November 2026 13 min read
On this page

DeepSeek V4.1, V4 and R1 are three generations of DeepSeek's open-weight Mixture of Experts language models: R1 was DeepSeek's first reasoning model in January 2025, V4 arrived as a preview on 24 April 2026, and V4.1 Flash followed on 10 September 2026. This page compares them side by side: API names, release dates, status, parameter counts, context window, output limits, prices, thinking modes, licences, provider-reported benchmark figures and retirement dates. It is written for developers migrating from R1 or V4, buyers estimating cost and architects who need to know which DeepSeek model name is still callable. Models are listed by release date, newest first, an objective order that implies no ranking. Every fact comes from DeepSeek's API documentation, change log, Hugging Face model cards and privacy policy, cited in the sources list.

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

  • Two of the three are on the API today. DeepSeek's pricing page lists deepseek-flash (DeepSeek-V4.1-Flash-1017) and deepseek-v4-pro (DeepSeek-V4-Pro-0813); R1 is no longer served under any current API name and is available as open weights.
  • Context grew from 128K to 1M tokens. The R1 card lists a 128K context; V4 and V4.1 Flash list 1M tokens with output up to 384K on the API.
  • V4.1 Flash costs less than a quarter of V4-Pro. Cache-miss input is $0.30 against $1.32 per million tokens at peak, and output $1.20 against $3.96; off-peak rates are half.
  • All three use the MIT licence. Each model card states that the repository and model weights are licensed under the MIT License.
  • Reasoning moved from a separate model to a mode. R1 was a dedicated reasoning model behind deepseek-reasoner; V4 and V4.1 Flash offer thinking and non-thinking modes in one model, with thinking on by default.

What is DeepSeek V4.1 vs V4 vs R1?

DeepSeek publishes general models (the V series) and, until 2025, a separate reasoning line (R1). The change log shows the two lines merging: on 21 August 2025 DeepSeek-V3.1 introduced hybrid reasoning, with one model supporting thinking and non-thinking modes, and deepseek-reasoner moved from R1 to V3.1's thinking mode. V4 and V4.1 continue that design.

  • DeepSeek V4.1 (Flash): described in the change log as the smallest model in DeepSeek's new architecture, with native multimodal visual understanding. It is served as deepseek-flash. See the DeepSeek V4.1 Flash profile.
  • DeepSeek V4 (Pro and Flash): two text models released as a preview on 24 April 2026. V4-Pro is current; V4-Flash was retired on 10 September 2026. See the DeepSeek V4 (Pro and Flash) profile.
  • DeepSeek R1: a reasoning model trained from DeepSeek-V3-Base with reinforcement learning, released with its paper in January 2025 and updated as R1-0528 in May 2025. Six distilled models built on Qwen and Llama bases were published alongside it.

For the full DeepSeek line-up, see DeepSeek Models.

Every model at a glance

As of 2026-11-15. Source: DeepSeek pricing page, change log and Hugging Face model cards. "Not stated by DeepSeek" means not stated on the pages checked (2026-11-15).

FieldV4.1 FlashV4 (Pro / Flash)R1
API namedeepseek-flashdeepseek-v4-pro; deepseek-v4-flash (retired, routed to V4.1 Flash)None current; formerly deepseek-reasoner
Current versionDeepSeek-V4.1-Flash-1017DeepSeek-V4-Pro-0813DeepSeek-R1-0528 (latest weights)
StatusCurrentPro current; Flash retired 2026-09-10Open weights only
Total / active parameters763B total (card); active Not stated by DeepSeek on the 1017 cardPro 1.6T / 49B; Flash 284B / 13B (preview card)671B / 37B (R1 card)
Context window1M1M128K
Max output (API)384K384KNot stated by DeepSeek for the API; evaluations used 32,768 (R1) and 64K (R1-0528)
InputText, imageText (vision not supported on deepseek-v4-pro)Text
ThinkingThinking and non-thinking; thinking defaultThinking and non-thinking; thinking default; effort low, high, maxAlways reasons
LicenceMITMITMIT
Knowledge cutoffNot stated by DeepSeekNot stated by DeepSeekNot stated by DeepSeek

The Hugging Face header for R1 shows 684B, above the card text figure, and the R1-0528 card shows 685B. Use the card text for architecture and the size box for storage planning.

When was each model released?

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

DateEventSource
2026-11-13Web search, file input, URL reading, model safety checks and other application APIs enabled for deepseek-flashChange log
2026-10-17V4.1 Flash upgraded to DeepSeek-V4.1-Flash-1017; calling method and pricing unchangedChange log
2026-09-10DeepSeek-V4.1-Flash released as deepseek-flash; V4-Flash and V4-Flash-Vision-Exp retired and routed; API prices reducedChange log
2026-08-13V4-Pro generally available; thinking effort low, high or max; peak and off-peak pricing from 16 AugustChange log
2026-07-31V4-Flash public beta (V4-Flash-0731, re-post-trained on the preview architecture)Change log
2026-07-24Stated discontinuation date for deepseek-chat and deepseek-reasonerChange log
2026-04-24DeepSeek-V4 preview: V4-Pro and V4-Flash on the APIChange log
2025-08-21DeepSeek-V3.1: deepseek-reasoner moves from R1 to V3.1 thinking modeChange log
2025-05-28deepseek-reasoner upgraded to DeepSeek-R1-0528Change log; R1-0528 card
2025-01-20DeepSeek-R1 served as deepseek-reasonerChange log

Between R1 and V4, DeepSeek also released V3-0324, V3.1-Terminus, V3.2-Exp and V3.2 under the deepseek-chat and deepseek-reasoner names; those are outside the scope of this page.

How much does each model cost?

As of 2026-11-15. Price per million tokens on the DeepSeek API, shown with a dollar sign on the pricing page. Peak is 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday, excluding Chinese public holidays; all other hours are off-peak at half the peak rate.

ModelCache-hit input (peak / off-peak)Cache-miss input (peak / off-peak)Output (peak / off-peak)Source
V4.1 Flash (deepseek-flash)$0.006 / $0.003$0.30 / $0.15$1.20 / $0.60Pricing page
V4-Pro (deepseek-v4-pro)$0.044 / $0.022$1.32 / $0.66$3.96 / $1.98Pricing page
V4-Flash (deepseek-v4-flash)Retired; requests are served by V4.1-Flash-1017 and billed at the Flash pricePricing page
R1Not stated by DeepSeek on the current pricing page; self-hosting cost depends on your infrastructurePricing page

Batch pricing is not stated by DeepSeek on the pages checked. For other providers, see the LLM Pricing Tracker.

What changed between models?

  • R1 to R1-0528: the R1-0528 card calls it a minor version upgrade with deeper reasoning, a reduced hallucination rate, enhanced function calling, system prompt support and no need for the <think> prefix. The change log adds JSON output and function calling on the API.
  • R1 to V4: one model now covers thinking and non-thinking work (inherited from V3.1), context rises from 128K to 1M, and the API adds the Anthropic-compatible interface alongside the OpenAI format. The V4 preview card describes a hybrid of Compressed Sparse Attention and Heavily Compressed Attention, FP4 and FP8 mixed precision, and three modes: Non-think, Think High and Think Max.
  • API names: from 24 April to 24 July 2026, deepseek-chat and deepseek-reasoner pointed to V4-Flash non-thinking and thinking modes, so code written for R1's name ran on V4-Flash during the transition without a change.
  • Pricing model: peak and off-peak billing, with off-peak at half the peak price, took effect at 16:00 UTC on 16 August 2026 with V4-Pro general availability; R1-era pricing is no longer published.
  • V4 to V4.1 Flash: image input arrives natively, the API name changes to deepseek-flash, and prices fell on 10 September 2026. The 1017 update aims at a better cost-performance trade-off and shorter task execution, per the change log.
  • Usage settings: the R1 card recommends temperature 0.5 to 0.7 (0.6) and no system prompt; the V4 and V4.1-Flash-1017 cards recommend temperature 1.0 and top_p 1.0, and the 1017 card max_tokens of at least 256K.

What workloads does DeepSeek position each model for?

ModelPositioning (provider wording, paraphrased)Source
V4.1 FlashSmallest model in the new architecture, with native multimodal visual understanding; 1017 update improves reasoning, coding, agentic and visual tasksChange log; 1017 card
V4-ProLarger V4 model; V4-Flash in maximum reasoning mode trails Pro on pure knowledge tasks and the most complex agentic workflowsV4 preview card
R1Reasoning on maths, code and reasoning tasks; supports commercial use and distillationR1 card

What is the context window of each model?

ModelContextMax outputNotesSource
V4.1 Flash1M384KConcurrency limit 2,500Pricing page
V4-Pro1M384KConcurrency limit 500; for local Think Max, the card recommends at least a 384K context windowPricing page; V4 preview card
R1128KNot stated by DeepSeek for the APIEvaluation generation length 32,768 (R1) and 64K (R1-0528)R1 and R1-0528 cards

Capabilities and tools compared

CapabilityV4.1 FlashV4-ProR1
JSON outputYesYesAdded with R1-0528 on the API (change log)
Tool callsYesYesFunction calling added with R1-0528 (change log)
FIM completion (Beta)Non-thinking mode onlyNon-thinking mode onlyNot stated by DeepSeek
VisionYesNoNo (text model)
Responses and Anthropic APIsYesYesNot stated by DeepSeek
Web search, file input, URL readingYes, from 2026-11-13Not stated by DeepSeekNot stated by DeepSeek

Source: DeepSeek pricing page and change log, as of 2026-11-15.

Where is each model available?

V4.1 Flash and V4-Pro are on the DeepSeek API at https://api.deepseek.com (OpenAI format) and https://api.deepseek.com/anthropic (Anthropic format). Weights for all three generations are on Hugging Face under the deepseek-ai organisation: DeepSeek-V4.1-Flash-1017, DeepSeek-V4-Pro and DeepSeek-V4-Pro-0813, and DeepSeek-R1, DeepSeek-R1-0528 and the R1 distilled models (1.5B, 7B, 8B, 14B, 32B and 70B, plus R1-0528-Qwen3-8B). Third-party cloud availability is outside the scope of this page. See Self-Hosted LLMs vs Hosted LLM APIs.

R1 distilled models

Distilled modelBase modelSource
DeepSeek-R1-Distill 1.5BQwen2.5-Math-1.5BR1 card
DeepSeek-R1-Distill 7BQwen2.5-Math-7BR1 card
DeepSeek-R1-Distill 8BLlama-3.1-8BR1 card
DeepSeek-R1-Distill 14BQwen2.5-14BR1 card
DeepSeek-R1-Distill 32BQwen2.5-32BR1 card
DeepSeek-R1-Distill 70BLlama-3.3-70B-InstructR1 card
DeepSeek-R1-0528-Qwen3-8BQwen3 8B Base (chain-of-thought distilled from R1-0528)R1-0528 card

The distilled models inherit their base model's architecture, so the base model's own licence terms sit alongside DeepSeek's MIT terms; check both before commercial use. See the Open-Weight Model Licence Tracker. V4 and V4.1 have no distilled releases listed on the cards checked.

Model cards and system cards

Each model has a Hugging Face card; DeepSeek publishes no separate system card on the pages checked (2026-11-15). The R1 card describes R1-Zero (reinforcement learning without supervised fine-tuning) and R1 (cold-start data, two RL and two SFT stages). The V4 preview card covers architecture, precision and benchmark tables. The V4.1-Flash-1017 card calls the release a minor update and cites a technical report titled "Pushing the Limits of KV Cache Compression".

Benchmarks

Provider-reported figures as shown on the cards and change log checked on 2026-11-15. Settings differ between cards (V4 figures are the Max modes; V4.1 HLE is without tools), so compare within a row with care. Independent leaderboards were not checked and are not shown.

BenchmarkV4.1-Flash-1017V4-Pro MaxV4-Flash MaxR1R1-0528
Humanity's Last Exam (HLE)42.6 (w/o tools)37.734.88.517.7
GPQA DiamondNot stated by DeepSeek90.188.171.581.0
SWE-bench VerifiedNot stated by DeepSeek80.679.049.257.6
MMLU-ProNot stated by DeepSeek87.586.284.085.0
Terminal-Bench42.3 (version 4.0)67.9 (version 2.0)56.9 (version 2.0)Not stated by DeepSeekNot stated by DeepSeek

Sources: V4.1-Flash-1017 card and change log; V4 preview card; R1-0528 card (R1 and R1-0528 columns). The R1 card separately lists AIME 2024 79.8, MATH-500 97.3 and a Codeforces rating of 2029. Terminal-Bench versions differ, so those figures are not comparable with each other.

Other provider-reported figures by model:

  • V4.1-Flash-1017: DeepSWE v1.1 77.0, NL2Repo-Bench 68.2, Automation-Bench v1.0.6 60.5, Apex Bench v1.1 72.1 and, for vision with tools, Chartography 83.0 and BabyVision 92.5.
  • V4-Pro Max: LiveCodeBench 93.5, HMMT 2026 Feb 95.2 and a Codeforces rating of 3206 (V4-Flash Max: 91.6, 94.8 and 3052).
  • R1-0528: AIME 2025 87.5 (R1: 70.0) and, per the change log, Aider 71.6 (R1: 57.0) and Tau-bench 53.5 Airline and 63.9 Retail.

Deprecations and retirements

Model or itemDetailsSource
deepseek-flashCurrent; always points to the latest Flash versionPricing page; change log
deepseek-v4-proCurrent; DeepSeek stated service continues after 14 September 2026 with unchanged billingChange log
deepseek-v4-flash, deepseek-v4-flash-vision-expRetired 2026-09-10; names still accepted and routed to V4.1-Flash-1017Pricing page; change log
deepseek-reasoner (former R1 name), deepseek-chatScheduled for discontinuation on 2026-07-24, announced 2026-04-24Change log
R1 weightsStill published on Hugging Face under MITR1 card

Data handling and security

DeepSeek's privacy policy (last updated 10 February 2026) states that DeepSeek directly collects, processes and stores personal data in the People's Republic of China, lists model training and improvement among its purposes for user input, and lists a right to opt out of training use by request to DeepSeek. For Open Platform customers it covers payment data, while end-user data collected by developers' own apps is the developer's responsibility. Self-hosting the MIT weights keeps prompts on your own infrastructure. Before adoption, see the DeepSeek and High-Risk AI Vendor Assessment Guide, the DeepSeek Federal Exclusion Orders and FASCSA Guide and the Open-Weight Model Security Guide.

How to choose a model: factual criteria

  • Image input: only V4.1 Flash accepts images on the API.
  • Hosted API: V4.1 Flash and V4-Pro are callable; R1 requires self-hosting or a third party.
  • Price: V4.1 Flash list prices are lower than V4-Pro at both peak and off-peak.
  • Self-hosting footprint: R1 is 671B total and 37B active; V4-Pro is 1.6T and 49B; V4.1 Flash is 763B on its card. The R1 distilled models run from 1.5B to 70B.
  • Version stability: deepseek-flash moves to new Flash weights without a code change; pin a version by self-hosting a dated checkpoint.

Practitioner checklist

  • Replace any remaining deepseek-reasoner or deepseek-chat calls; both were scheduled for discontinuation on 24 July 2026.
  • Move deepseek-v4-flash calls to deepseek-flash; the old name is routed only as a legacy convenience.
  • Remove R1-era settings (no system prompt, forced <think> prefix, temperature 0.6) when moving to V4 or V4.1, and retest with the card's recommended settings.
  • Model cost with peak and off-peak rates in UTC and with thinking output included.
  • Run your own evaluation set; card scores use different settings and benchmark versions.
  • Complete a vendor risk assessment covering data location and training use before sending business data to the API.
  • Store DeepSeek API keys in a secrets manager and rotate them; see AI model provider API keys.
  • Watch the DeepSeek change log for retirements and routing changes.

Standards and references

Related NHI Mgmt Group resources: DeepSeek Models, DeepSeek V4.1 Flash, DeepSeek V4 (Pro and Flash), DeepSeek vs Qwen vs Kimi vs GLM, Frontier AI Models Compared.

PDF for subscribers

Take this guide with you

Every section, table and checklist in one PDF, with links to the sources and related guides. Subscribe to our weekly NHI & AI Identity Journal and we’ll email you a download link.

Free weekly newsletter

Subscribe to the NHI & AI Identity Journal

The latest on NHI and Agentic AI security – articles, research, breaches, news and events every week.

Bonus 33% off our NHI Course when you subscribe.

NHIMG Editorial Note
Researched and written by the NHI Mgmt Group editorial team using AI tools, under the editorial direction of Lalit Choda. Facts are drawn from the sources listed on this page and checked by our automated quality process. Last updated 15 November 2026. Spotted an error? Use our contact page.
NHI Mgmt Group, the independent authority on Non-Human Identity, IAM and Agentic AI security. nhimg.org

PDF Downloads for Newsletter Subscribers

Get the guide

Enter your details and we’ll email you a download link.

The full guide is free to read online. The PDF is a thank you for subscribing.

Fill in all fields marked * and tick the PDF box to continue.