DeepSeek V4-Flash Explained: From V4 Preview to V4.1-Flash-1017

DeepSeek V4-Flash began in April 2026 as the smaller, cheaper half of the DeepSeek V4 Preview. Six months on, the name covers a whole line of releases. The current API model is called deepseek-flash, and since 17 October 2026 it has been served by DeepSeek-V4.1-Flash-1017. DeepSeek describes that model as part of a new architecture family rather than a V4 build. This guide explains the Flash line as it stands on 23 October 2026: which model answers your requests, how each release got there, what it costs at peak and off-peak, how to send images, and where the open weights live. Every fact comes from DeepSeek’s own documentation, Change Log, release notes and Hugging Face model cards; benchmark figures are DeepSeek-reported and dated.
DeepSeek V4-Flash today: deepseek-flash and V4.1-Flash-1017
The legacy model name deepseek-v4-flash still works, but it no longer means a V4 model. DeepSeek’s Models & Pricing page lists deepseek-flash as the model name and DeepSeek-V4.1-Flash-1017 as the model version behind it. Requests sent under the older Flash names are served by the same model and billed at the Flash price.
| Attribute | deepseek-flash (as of 23 Oct 2026) |
|---|---|
| Model version | DeepSeek-V4.1-Flash-1017 (released 17 Oct 2026) |
| Legacy names still accepted | deepseek-v4-flash, deepseek-v4-flash-vision-exp (served by V4.1-Flash-1017, Flash price) |
| Base URL, OpenAI format | https://api.deepseek.com |
| Base URL, Anthropic format | https://api.deepseek.com/anthropic |
| Thinking mode | Non-thinking and thinking; thinking is the default (effort high) |
| Context length | 1M tokens |
| Maximum output | 384K tokens |
| Vision (image input) | Supported |
| Other features | JSON Output, Tool Calls, Responses API, Anthropic API, Chat Prefix Completion (beta), FIM Completion (beta, non-thinking mode only) |
| Concurrency limit | 2,500 per account |
| Price per 1M tokens, peak / off-peak | Cache hit $0.006 / $0.003; cache miss $0.30 / $0.15; output $1.20 / $0.60 |
| Open weights | MIT licence, deepseek-ai/DeepSeek-V4.1-Flash-1017 on Hugging Face |
Source: DeepSeek Models & Pricing and Rate Limit & Isolation, accessed 23 October 2026. For the family as a whole, see the DeepSeek V4 overview; the larger sibling is covered on the DeepSeek V4-Pro page.
Flash release history, April to October 2026
Five Flash releases appear on DeepSeek’s official pages. The first two share one architecture; the last two belong to the new V4.1 family.
| Date | Release | API model name | What DeepSeek said |
|---|---|---|---|
| 24 Apr 2026 | DeepSeek-V4-Flash (preview) | deepseek-v4-flash |
Preview MoE model, 284B total / 13B active parameters, 1M context |
| 31 Jul 2026 | DeepSeek-V4-Flash-0731 | deepseek-v4-flash |
Official V4-Flash in API public beta; same architecture and size as the preview, re-post-trained |
| 21 Aug 2026 | DeepSeek-V4-Flash-Vision-Exp | deepseek-v4-flash-vision-exp |
Experimental multimodal (vision) model |
| 10 Sep 2026 | DeepSeek-V4.1-Flash | deepseek-flash |
Smallest model in a new architecture family, native visual understanding; V4-Flash and V4-Flash-Vision-Exp retired; prices reduced |
| 17 Oct 2026 | DeepSeek-V4.1-Flash-1017 | deepseek-flash |
Upgrade of V4.1-Flash; API calling method and pricing unchanged |
Sources: DeepSeek API Change Log and the release notes linked below, accessed 23 October 2026.
24 April 2026: the V4-Flash preview
The DeepSeek-V4 Preview release note introduced two models, V4-Pro and V4-Flash, both with a 1M-token context and both thinking and non-thinking modes. The preview’s V4-Flash had 284B total parameters with 13B active. Those figures describe the April preview only; DeepSeek has not published them for the current Flash model. DeepSeek said the preview Flash’s reasoning capabilities closely approached V4-Pro’s and that it performed on par with V4-Pro on simple agent tasks. The preview’s Hugging Face card described a hybrid attention design combining Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA).
On the same day, the two older model names were pointed at the preview: deepseek-chat went to deepseek-v4-flash in non-thinking mode and deepseek-reasoner to its thinking mode. DeepSeek announced that both names would be fully retired after 24 July 2026, 15:59 UTC, and they no longer appear in the official documentation.
31 July 2026: DeepSeek-V4-Flash-0731
The official V4-Flash entered API public beta under the unchanged name deepseek-v4-flash. According to the Change Log, DeepSeek-V4-Flash-0731 “keeps the same model architecture and size as DeepSeek-V4-Flash-Preview, and was only re-post-trained.” Its model card adds that it carries a speculative-decoding module, the same structure as DeepSeek-V4-Flash-DSpark. The release also brought native support for the Responses API format. On 13 August 2026, alongside the V4-Pro GA release, DeepSeek added three thinking-effort levels (low, high, max) for both V4-Pro and V4-Flash and announced peak and off-peak pricing from 16:00 UTC on 16 August 2026.
21 August 2026: DeepSeek-V4-Flash-Vision-Exp
DeepSeek released an experimental multimodal model, reachable as deepseek-v4-flash-vision-exp. Its model card says it builds on the DeepSeek-V4-Flash architecture with added visual modules and continued training. DeepSeek said its text capabilities were on par with the official V4-Flash, with a significant gain on agent benchmarks that need visual understanding.
10 September 2026: DeepSeek-V4.1-Flash
The 10 September 2026 release note introduced DeepSeek-V4.1-Flash as “the smallest model in our new architecture family, with native visual understanding.” For that release, DeepSeek gave a 552B-parameter MoE with a Causal Encoder–Decoder design that activates 8B parameters for input and 16B for output. It also said V4.1-Flash’s KV cache needs a quarter of the HBM and an eighth of the SSD storage of the previous generation.
The API model name became deepseek-flash. V4-Flash and V4-Flash-Vision-Exp were retired, and their names were “temporarily” routed to V4.1-Flash for compatibility. Prices were reduced from 04:00 UTC on 10 September 2026, with the peak/off-peak structure kept. The same note announced that deepseek-v4-pro requests would be routed to V4.1-Flash from 14 September. The Change Log entry for that date records the opposite decision: DeepSeek chose to keep serving V4-Pro after 14 September 2026 with billing unchanged, and the pricing page still lists DeepSeek-V4-Pro-0813 today.
17 October 2026: DeepSeek-V4.1-Flash-1017
The 17 October 2026 release note upgraded deepseek-flash to DeepSeek-V4.1-Flash-1017 on the web, the app and the API. DeepSeek reported stronger benchmark results, faster task completion and shorter outputs than the September build; the API calling method and pricing stayed the same. The model card calls it a minor update to DeepSeek-V4.1-Flash and publishes no separate parameter counts, so the 552B / 8B / 16B figures above are only stated for the 10 September release. The same day, DeepSeek opened its Search and Fetch APIs to all users at $0.75 per 1,000 successful requests each, billed per request rather than per token.
DeepSeek-reported benchmarks
The site runs no evaluations of its own. All scores below are DeepSeek’s, taken from the Change Log and model cards with the date each was published. DeepSeek ran them with its own harness settings, which differ between benchmarks.
Terminal-Bench 2.1 across the Flash line
| Model | Terminal-Bench 2.1 | Source (date) |
|---|---|---|
| DeepSeek-V4-Flash (preview) | 61.8 | V4-Flash-0731 model card (31 Jul 2026) |
| DeepSeek-V4-Flash-0731 | 82.7 | Change Log (31 Jul 2026) |
| DeepSeek-V4-Flash-Vision-Exp | 83.9 | Change Log (21 Aug 2026) |
| DeepSeek-V4.1-Flash | 90.6 | Change Log (10 Sep 2026) |
| DeepSeek-V4-Pro-0813 (for reference) | 87.9 | Change Log (13 Aug 2026) |
DeepSeek has not published a Terminal-Bench 2.1 score for V4.1-Flash-1017. On 10 September 2026, DeepSeek also claimed that V4.1-Flash delivers “benchmark results ahead of flagship models, including DeepSeek-V4-Pro.” That is a vendor claim.
V4.1-Flash-1017 against the September V4.1-Flash
| Benchmark | V4.1-Flash-1017 | V4.1-Flash (Sep 2026) |
|---|---|---|
| HLE (without tools) | 42.6 (44.0 on the pure-text subset) | 36.8 (39.1 on the pure-text subset) |
| Terminal-Bench 4.0 | 42.3 | 31.2 |
| DeepSWE v1.1 | 77.0 | 74.2 |
| ProgramBench | 29.1 | 20.3 |
| Automation-Bench | 60.5 (v1.0.6) | 54.8 |
| Agents’ Last Exam | 35.4 | 31.8 |
| Chartography (with tools) | 83.0 | 78.9 |
| BabyVision (with tools) | 92.5 | 89.6 |
| ZeroBench-main (with tools) | 51.0 | 49.0 |
Source: DeepSeek API Change Log, entries of 17 October 2026 (V4.1-Flash-1017) and 10 September 2026 (V4.1-Flash); the same figures appear in the V4.1-Flash-1017 model card. In the same card, DeepSeek’s comparison table puts Opus 5 ahead of V4.1-Flash-1017 on HLE without tools (56.3 against 42.6) and Terminal-Bench 4.0 (51.8 against 42.3), while V4.1-Flash-1017 has the higher DeepSWE v1.1 figure (77.0 against 74.0). These are DeepSeek’s figures for all models listed, not independent results. For a wider view of this year’s releases, see the 2026 DeepSeek benchmarks roundup.
Legacy model names and what they route to
deepseek-v4-flashis still accepted. The V4-Flash model behind it was retired on 10 September 2026; requests are served by DeepSeek-V4.1-Flash-1017 and billed at the Flash price.deepseek-v4-flash-vision-expis still accepted on the same terms: served by V4.1-Flash-1017, billed at the Flash price.deepseek-chatanddeepseek-reasonerwere the general-purpose names before V4. From 24 April 2026 they pointed todeepseek-v4-flashin non-thinking and thinking mode, and they were scheduled for retirement on 24 July 2026 at 15:59 UTC. They are no longer listed in DeepSeek’s documentation.
DeepSeek described the routing of the two V4-Flash names as temporary when it set it up on 10 September 2026. The safe migration is a one-word change: set model to deepseek-flash. Thinking mode is now on by default, so code that relied on deepseek-chat’s non-thinking behaviour should pass {"thinking": {"type": "disabled"}} explicitly.
deepseek-flash pricing, peak and off-peak

DeepSeek has charged peak and off-peak rates since 16:00 UTC on 16 August 2026. Off-peak rates are half the peak rates. Peak hours are 01:00–04:00 and 06:00–10:00 UTC, Monday to Friday, excluding Chinese public holidays. All other hours are off-peak, including weekends and Chinese public holidays in full. That is 35 of the week’s 168 hours (about 20.8%) at peak in a week without holidays.
| USD per 1M tokens | deepseek-flash peak | deepseek-flash off-peak | deepseek-v4-pro peak | deepseek-v4-pro off-peak |
|---|---|---|---|---|
| Input, cache hit | $0.006 | $0.003 | $0.044 | $0.022 |
| Input, cache miss | $0.30 | $0.15 | $1.32 | $0.66 |
| Output | $1.20 | $0.60 | $3.96 | $1.98 |
Source: DeepSeek Models & Pricing, accessed 23 October 2026. On Flash, a cache hit costs one fiftieth of a cache miss ($0.006 against $0.30 at peak), so repeated prompt prefixes matter; see the context caching guide. The full rate card, including V4-Pro, is on the DeepSeek API pricing page.
Worked example: 1,000,000 requests
Assume 1,000,000 requests, each with a 2,000-token system prompt served from cache, a 200-token user message that misses the cache, and a 300-token reply. That is 2,000,000,000 cached input tokens, 200,000,000 uncached input tokens and 300,000,000 output tokens.
deepseek-flash, peak
Cache-hit input: 2,000M tokens × $0.006/M = $12.00
Cache-miss input: 200M tokens × $0.30/M = $60.00
Output: 300M tokens × $1.20/M = $360.00
Total $432.00
deepseek-flash, off-peak
Cache-hit input: 2,000M tokens × $0.003/M = $6.00
Cache-miss input: 200M tokens × $0.15/M = $30.00
Output: 300M tokens × $0.60/M = $180.00
Total $216.00
The same workload on deepseek-v4-pro comes to $88.00 + $264.00 + $1,188.00 = $1,540.00 at peak and $44.00 + $132.00 + $594.00 = $770.00 off-peak, about 3.56 times the Flash total in both periods. To model your own mix of tokens and hours, use the DeepSeek pricing calculator.
Calling deepseek-flash from code
The API is OpenAI- and Anthropic-compatible. This Python example follows DeepSeek’s “Your First API Call” page and reads the key from the DEEPSEEK_API_KEY environment variable:
# pip3 install openai
import os
from openai import OpenAI
client = OpenAI(
api_key=os.environ.get("DEEPSEEK_API_KEY"),
base_url="https://api.deepseek.com",
)
response = client.chat.completions.create(
model="deepseek-flash",
messages=[
{"role": "system", "content": "You are a helpful assistant."},
{"role": "user", "content": "Explain peak and off-peak API pricing in two sentences."},
],
stream=False,
reasoning_effort="high",
extra_body={"thinking": {"type": "enabled"}},
)
print(response.choices[0].message.reasoning_content) # chain of thought
print(response.choices[0].message.content) # final answer
The same request with curl:
curl https://api.deepseek.com/chat/completions \
-H "Content-Type: application/json" \
-H "Authorization: Bearer ${DEEPSEEK_API_KEY}" \
-d '{
"model": "deepseek-flash",
"messages": [
{"role": "system", "content": "You are a helpful assistant."},
{"role": "user", "content": "Hello!"}
],
"thinking": {"type": "enabled"},
"reasoning_effort": "high",
"stream": false
}'
Points from DeepSeek’s Thinking Mode guide worth knowing:
- Thinking is on by default with effort
high. Setreasoning_efforttolow,highormax, or pass{"thinking": {"type": "disabled"}}for non-thinking replies. - In thinking mode,
temperature,presence_penaltyandfrequency_penaltyhave no effect. Ondeepseek-flash,top_ponly applies in thinking mode, within 0.95–1.0; in non-thinking mode it is fixed at 1.0. - The chain of thought comes back in
reasoning_content. When a request carriestools, earlier turns’reasoning_contentmust be passed back, or the API returns a 400 error.
The concurrency limit for deepseek-flash is 2,500 requests per account, counted from when a request is sent until its response completes, across all of the account’s API keys. Requests above the limit receive HTTP 429; DeepSeek accepts capacity-expansion requests at no extra cost. More detail is in the rate limits guide, and further patterns are in the API code examples. Start with getting an API key if you do not have one; JSON responses are covered in the JSON mode guide.
Sending images to deepseek-flash

Vision is supported on deepseek-flash and not on deepseek-v4-pro. According to DeepSeek’s Vision guide, the model can describe pictures, read text from screenshots and analyse charts. Images go in the content array of a user message, as a base64 data URL, a public http(s) URL, or a file_id from the Files API. The same methods work in the Responses API (input_image parts) and the Anthropic-format endpoint (image blocks).
import base64
import os
from openai import OpenAI
client = OpenAI(
api_key=os.environ.get("DEEPSEEK_API_KEY"),
base_url="https://api.deepseek.com",
)
with open("image.jpg", "rb") as f:
b64 = base64.b64encode(f.read()).decode("utf-8")
response = client.chat.completions.create(
model="deepseek-flash",
messages=[
{
"role": "user",
"content": [
{"type": "text", "text": "What is in this image?"},
{
"type": "image_url",
"image_url": {"url": f"data:image/jpeg;base64,{b64}"},
},
],
},
],
)
print(response.choices[0].message.content)
- Formats: JPEG, PNG, GIF and WebP, detected from the file content.
- Where: user messages only; images in system or assistant messages return a 400 error.
- Size limits: 48 MiB request body; 32 MiB per image sent as base64 or URL; 64 MiB per image through the Files API; up to 600 images per request; 8,192 px per side, or 4,096 px when a request has 15 or more images.
- Detail:
"detail": "low"downscales an image to 512×512 before inference. - Tokens: each image is resized and counted as at most 1,024 tokens, billed together with the text tokens.
At the 1,024-token maximum, 1,000 images add up to 1,024,000 tokens. At the cache-miss input rate that is 1.024 × $0.30 = $0.3072 at peak, or 1.024 × $0.15 = $0.1536 off-peak.
Flash or V4-Pro?

Both models share the 1M-token context, the 384K-token output limit, thinking mode and the same set of API formats. The differences that are documented:
- Price: V4-Pro costs 3.3 times as much as Flash for output, 4.4 times for cache-miss input and about 7.33 times for cache-hit input, at both peak and off-peak.
- Concurrency: 2,500 for
deepseek-flashagainst 500 fordeepseek-v4-pro. - Vision: Flash only.
- Model version:
deepseek-v4-prois DeepSeek-V4-Pro-0813, released on 13 August 2026. The model behind the Flash name has changed twice since then, on 10 September and 17 October 2026.
On DeepSeek’s own Terminal-Bench 2.1 figures, V4.1-Flash (90.6, 10 September 2026) is ahead of V4-Pro-0813 (87.9, 13 August 2026). Treat this as a vendor comparison rather than a verdict. The DeepSeek V4-Pro page covers the Pro model in full.
Open weights on Hugging Face
DeepSeek-V4.1-Flash-1017 is published at huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash-1017, and the 17 October release note also lists ModelScope. The card states that the repository and the model weights are licensed under the MIT License. It ships no Jinja chat template; prompt encoding is provided as a Python reference implementation in an encoding folder, with local inference instructions in an inference folder. DeepSeek recommends temperature 1.0, top_p 1.0, a 1M-token context window and max_tokens of at least 256K for local runs.
The earlier Flash checkpoints (the April preview, V4-Flash-0731, V4-Flash-Vision-Exp and the September V4.1-Flash) also remain on Hugging Face under MIT. For setting up a local runtime, see how to install DeepSeek locally.
Where Flash fits
- Coding and agent work: DeepSeek optimised V4.1-Flash-1017 for DeepSeek Harness, Tencent CodeBuddy and WorkBuddy, and the 10 September note names OpenCode as a partner. See DeepSeek for coding.
- Image understanding: screenshots, charts and photos in user messages, which V4-Pro does not accept.
- Long inputs: the 1M-token context applies to Flash as to Pro; the context length checker helps size a prompt.
- High-volume text work: writing and translation pipelines where output price dominates. See DeepSeek for writing and DeepSeek for translation.
For comparisons outside DeepSeek, see DeepSeek vs ChatGPT, DeepSeek vs Claude and the DeepSeek alternatives hub. The rest of the family is on the DeepSeek models hub.
Frequently asked questions
What is DeepSeek V4-Flash now?
It is the name of DeepSeek’s cheaper model line. The API model to use is deepseek-flash, which DeepSeek has served with DeepSeek-V4.1-Flash-1017 since 17 October 2026. V4.1-Flash is part of a new architecture family introduced on 10 September 2026; the original V4-Flash, first released as a preview on 24 April 2026, has been retired. Background on the family is on the DeepSeek V4 page.
Does the model name deepseek-v4-flash still work?
Yes. DeepSeek still accepts deepseek-v4-flash and deepseek-v4-flash-vision-exp, but the models behind them are retired. Requests are served by DeepSeek-V4.1-Flash-1017 and billed at the Flash price. DeepSeek called this routing temporary, so new code should use deepseek-flash.
How much does deepseek-flash cost?
Per 1M tokens as of 23 October 2026: cache-hit input $0.006 at peak and $0.003 off-peak, cache-miss input $0.30 peak and $0.15 off-peak, output $1.20 peak and $0.60 off-peak. Peak hours are 01:00–04:00 and 06:00–10:00 UTC on weekdays, excluding Chinese public holidays. See the DeepSeek API pricing page or the pricing calculator.
Can DeepSeek V4-Flash read images?
Yes. deepseek-flash accepts JPEG, PNG, GIF and WebP images in user messages, sent as base64, a public URL or a Files API file_id. Each image counts as at most 1,024 tokens. deepseek-v4-pro does not support vision.
How many parameters does V4-Flash have?
It depends on the release. The April 2026 V4-Flash preview had 284B total and 13B active parameters, and V4-Flash-0731 kept that architecture and size. DeepSeek’s 10 September 2026 note gives V4.1-Flash as a 552B-parameter MoE with 8B active parameters for input and 16B for output. DeepSeek has published no separate size figures for V4.1-Flash-1017, which its model card calls a minor update to V4.1-Flash.
How is thinking mode switched on or off?
Thinking is on by default. In the OpenAI Python SDK, pass extra_body={"thinking": {"type": "enabled"}} with reasoning_effort set to low, high or max; use {"type": "disabled"} to turn it off. The reasoning comes back in reasoning_content. More patterns are in the API code examples.
Are the V4.1-Flash-1017 weights open?
Yes. The weights are on Hugging Face at deepseek-ai/DeepSeek-V4.1-Flash-1017 under the MIT License, with a mirror on ModelScope. DeepSeek recommends temperature 1.0, top_p 1.0 and a max_tokens setting of at least 256K for local inference.
Sources & Methodology
Primary sources
- Official pricing pageDeepSeek API Docs: Models & PricingModel names and versions, legacy-name routing, features, vision support, context and output limits, peak/off-peak prices and hours, concurrency limits, Search & Fetch price (accessed 23 Oct 2026)Last checked: October 23, 2026
- Official changelogDeepSeek API Change LogRelease dates and descriptions of V4 Preview, V4-Flash-0731, V4-Pro GA, V4-Flash-Vision-Exp, V4.1-Flash and V4.1-Flash-1017; the 0731 same-architecture statement; V4-Pro kept after 14 Sep 2026; DeepSeek-reported benchmarks (accessed 23 Oct 2026)Last checked: October 23, 2026
- Official release noteDeepSeek news: DeepSeek-V4 Preview Release (24 Apr 2026)Preview parameter counts (284B / 13B), preview positioning, deepseek-chat / deepseek-reasoner routing and retirement date (accessed 23 Oct 2026)Last checked: October 23, 2026
- Official release noteDeepSeek news: DeepSeek-V4-Pro GA Release (13 Aug 2026)low/high/max effort levels and peak/off-peak pricing from 16:00 UTC on 16 Aug 2026 (accessed 23 Oct 2026)Last checked: October 23, 2026
- Official release noteDeepSeek news: DeepSeek-V4.1-Flash Release (10 Sep 2026)New architecture family, 552B MoE / 8B input / 16B output active parameters, KV-cache claims, retirement and temporary routing of V4-Flash names, price reduction from 04:00 UTC 10 Sep 2026, vendor claim of results ahead of V4-Pro, announced V4-Pro routing (accessed 23 Oct 2026)Last checked: October 23, 2026
- Official release noteDeepSeek news: DeepSeek-V4.1-Flash-1017 Release (17 Oct 2026)V4.1-Flash-1017 launch on web, app and API, harness optimisations, Search & Fetch opening, Hugging Face and ModelScope links (accessed 23 Oct 2026)Last checked: October 23, 2026
- Model cardHugging Face: deepseek-ai/DeepSeek-V4.1-Flash-1017MIT licence, 'minor update' description, benchmark table incl. competitor figures, recommended local sampling parameters, encoding folder (accessed 23 Oct 2026)Last checked: October 23, 2026
Technical references
- Official documentationDeepSeek API Docs: Your First API CallPython and curl sample code, base URLs, model names (accessed 23 Oct 2026)Last checked: October 23, 2026
- Official guideDeepSeek API Docs: Thinking ModeThinking toggle, default effort, effort levels, parameter behaviour (temperature, top_p), reasoning_content handling (accessed 23 Oct 2026)Last checked: October 23, 2026
- Official guideDeepSeek API Docs: VisionImage input methods, request format, formats, size limits, detail levels, image token counting (accessed 23 Oct 2026)Last checked: October 23, 2026
- Official documentationDeepSeek API Docs: Rate Limit & IsolationConcurrency limits 2,500 / 500, counting rules, HTTP 429, capacity expansion (accessed 23 Oct 2026)Last checked: October 23, 2026
- Model cardHugging Face: deepseek-ai/DeepSeek-V4.1-FlashCross-check of September V4.1-Flash benchmark figures and architecture description (accessed 23 Oct 2026)Last checked: October 23, 2026
- Model cardHugging Face: deepseek-ai/DeepSeek-V4-Flash-0731Speculative-decoding module (DSpark structure), preview Terminal Bench 2.1 figure 61.8 (accessed 23 Oct 2026)Last checked: October 23, 2026
- Model cardHugging Face: deepseek-ai/DeepSeek-V4-Flash-Vision-ExpVision-Exp built on the V4-Flash architecture with visual modules (accessed 23 Oct 2026)Last checked: October 23, 2026
- Model cardHugging Face: deepseek-ai/DeepSeek-V4-FlashPreview hybrid attention (CSA + HCA) description; MIT licence of earlier checkpoints (accessed 23 Oct 2026)Last checked: October 23, 2026
Methodology
This page is compiled from DeepSeek's official API documentation (Models & Pricing, Your First API Call, Thinking Mode, Vision, Rate Limit & Isolation), the API Change Log, DeepSeek's release notes of 24 Apr, 13 Aug, 10 Sep and 17 Oct 2026, and the Hugging Face model cards of the Flash releases, all accessed on 23 Oct 2026. DeepSeek AI Guide did not run, measure or benchmark any model; every benchmark figure is DeepSeek-reported and dated, and the cost examples are arithmetic on the published rates.
Data confidence
High for model names, routing, prices, limits and release dates, which come straight from DeepSeek's documentation as of 23 Oct 2026. Benchmark figures are vendor-reported and use DeepSeek's own harness settings, so they indicate direction rather than independent results.
Editorial note
Vendor-reported figures are not always independently replicated. Benchmarks at the frontier change quickly; expect this article to need a refresh whenever DeepSeek, OpenAI, Anthropic, or Google ship a new model.
