TL;DR: Chinese AI startup DeepSeek launched DeepSeek-V4.1-Flash on Thursday, Sept. 10, 2026, describing it as the smallest model in its new architecture family and putting it live on its API as deepseek-flash.

Launch confirms new efficiency-focused model

BEIJING — DeepSeek launched DeepSeek-V4.1-Flash on Thursday, Sept. 10, 2026, according to Reuters, which reported the release from Beijing on Sept. 10 with reporting by Shi Bu and Eduardo Baptista and editing by Thomas Derpinghaus.

The company said in its own announcement that V4.1-Flash is the smallest model in its new architecture family, with native visual understanding, according to the official DeepSeek announcement titled "Introducing DeepSeek-V4.1-Flash: smarter, faster, more efficient."

DeepSeek said the new model is designed for greater capability, faster inference, higher throughput and scaling to larger models, according to Reuters and the DeepSeek announcement.

The release came as the company is starting to prepare for an initial public offering on Shanghai's tech-focused STAR Market, as Reuters has reported, according to Reuters.

The Hindu BusinessLine, carrying Reuters reporting, also reported the Sept. 10 launch and the STAR Market IPO preparation, with an update timestamp of Sept. 10, 2026 at 12:00 PM.

What V4.1-Flash is

DeepSeek describes V4.1-Flash as a 552-billion-parameter mixture-of-experts model with native multimodal visual understanding, according to the DeepSeek announcement and Caixin Global reporting by Yang Zirui published Sept. 10, 2026 at 5:47 p.m. GMT+8.

The design uses what DeepSeek calls a Causal Encoder-Decoder architecture, with 8 billion active parameters for input and 16 billion for output, according to the DeepSeek announcement.

The Next Web, in reporting published Sept. 10, 2026 at 3:48 p.m. by Ana Maria Constantin, described the same totals and activation split and called V4.1-Flash a 552-billion-parameter open model.

DeepSeek said new pretraining methods plus larger-scale reinforcement-learning post-training deliver benchmark results ahead of flagship models, including DeepSeek-V4-Pro, according to the DeepSeek announcement.

The Next Web reported that DeepSeek announced the model in a thread on X and published the weights on Hugging Face under the MIT licence, which lets anyone download, modify and run it.

The same report said the model handles up to 1 million tokens of context and understands images natively, according to its model card, and said DeepSeek reports pre-training on 45 trillion tokens.

S5 Labs, in analysis published Sept. 10, 2026 by Joshua Wendt, described the model card as listing a 552B-backbone-parameter mixture-of-experts model with a one-million-token context window and the same 8B during input processing and 16B during generation split.

Why the asymmetric design targets agents

DeepSeek said the asymmetric architecture delivers more intelligence at less cost, with the small active-parameter counts for reading input and writing output, according to the DeepSeek announcement.

The Next Web reported that the split is aimed at AI agents, because an agent that calls tools repeatedly spends much of its time reading fresh input, so cheaper reading means cheaper agents.

A central engineering claim concerns memory for long sessions. Models keep a cache of what they have already processed, known as the KV cache, which avoids recomputing everything at every step, according to The Next Web.

DeepSeek said that compared with the previous generation, V4.1-Flash's KV cache needs just one-quarter the high-bandwidth memory and one-eighth the SSD storage, according to the DeepSeek announcement.

The company said cache-hit charges often account for a large share of agent costs and that compressing the cache cuts those costs significantly, according to the DeepSeek announcement.

The Next Web reported that DeepSeek says V4.1-Flash needs 890 bytes per token for that cache, about a quarter of what V4-Flash needed and roughly 437 times less than its first model in 2023.

S5 Labs reported the same 890 bytes per token figure as a global KV cache, roughly a quarter of V4 Flash's footprint, with persistent cache storage around one-eighth as large, and noted these are cache reductions, not equivalent reductions in memory needed for the whole model.

Live on the API

DeepSeek said V4.1-Flash is now live on the DeepSeek API with native multimodal support, with users setting the model to deepseek-flash, according to the DeepSeek announcement.

The company said V4-Flash and V4-Flash-Vision-Exp are retired and that for compatibility deepseek-v4-flash and deepseek-v4-flash-vision-exp temporarily route to V4.1-Flash, according to the DeepSeek announcement.

S5 Labs reported the same retirement and temporary routing, and reported the new API name as deepseek-flash, in its Sept. 10 analysis.

DeepSeek said official partners WorkBuddy, including CodeBuddy, and OpenCode now fully support V4.1-Flash, according to the DeepSeek announcement.

The Next Web also reported that coding tools WorkBuddy and OpenCode already support the model.

DeepSeek said it will work closely with the open-source community on V4.1-Flash inference support and explore more deployment options, according to the DeepSeek announcement.

The company invited operators planning a large-scale deployment with 2,000 GPUs plus a storage cluster to contact it, according to the DeepSeek announcement, a detail also noted by The Next Web.

V4-Pro phaseout plan announced for Sept. 14

DeepSeek said tests by multiple parties put V4.1-Flash ahead of V4-Pro on performance, cost, speed and total runtime, and said it is phasing out V4-Pro, according to the DeepSeek announcement.

Starting at 04:00 UTC on Sept. 14, 2026, all deepseek-v4-pro requests will route to V4.1-Flash at V4.1-Flash rates, continuing until V4.1-Pro launches, according to the DeepSeek announcement.

The Next Web reported the same Sept. 14, 04:00 UTC switchover, with those requests billed at the cheaper Flash rates until a V4.1-Pro arrives, and noted DeepSeek gave no date for that next release.

S5 Labs likewise reported that V4 Pro follows at 04:00 UTC on Sept. 14, 2026, when deepseek-v4-pro requests will receive V4.1-Flash at Flash prices until V4.1 Pro launches, and that DeepSeek had not supplied a date for that next release in the announcement.

The Next Web described the change as a behavior change even for applications whose requests continue to succeed, because outputs can change without a configuration change.

S5 Labs advised teams maintaining regression suites to record the migration boundary and rerun representative tasks, checking outputs as well as whether requests succeed.

Prices take effect Sept. 10

DeepSeek said the more efficient architecture lets it serve more users at lower cost and that it is passing the savings to customers, according to the DeepSeek announcement.

The company said peak and off-peak pricing continues to balance demand, with off-peak rates at 50% of peak rates, and that new pricing takes effect at 04:00 UTC on Sept. 10, 2026, according to the DeepSeek announcement.

S5 Labs listed the published dollar rates per million tokens as $0.003 for cached input off-peak and $0.006 at peak, $0.15 for uncached input off-peak and $0.30 at peak, and $0.60 for output off-peak and $1.20 at peak.

The same analysis listed peak hours as 01:00-04:00 and 06:00-10:00 UTC, Monday through Friday, with all other hours off-peak, and said the new rates took effect at 04:00 UTC on Sept. 10.

For arithmetic illustration, S5 Labs calculated one million uncached input tokens plus one million output tokens as $0.75 off-peak or $1.50 at peak rates, noting retries and longer generated answers can change the bill even when token prices fall.

The Next Web reported off-peak pricing of $0.60 per million tokens of output and $0.003 for cached input, with peak rates double and applying on weekday mornings UTC, and said the new rates took effect at 04:00 UTC on Thursday.

Bloomberg reported on Sept. 10, 2026 by Saritha Rai that DeepSeek rolled out an AI model that charges as little as a fraction of a cent per million tokens.

Benchmarks show coding strength, reasoning gaps

Caixin Global reported that DeepSeek says the model can outperform some flagship systems while cutting inference costs, as Chinese AI developers increasingly compete on both capability and price.

Bloomberg reported that the Chinese startup unveiled V4.1 Flash on Thursday as a slimmed-down platform it claims outperformed mainstays such as Moonshot's Kimi K3, yet offers a steep discount to competition.

The Next Web reported DeepSeek's own benchmark table puts V4.1-Flash level with the best closed models on some coding tests, with 74.2 on DeepSWE v1.1 against 74.0 for Anthropic's Claude Opus 5 and 73.0 for OpenAI's GPT-5.6 Sol.

On CyberGym, a cybersecurity test, the same table lists 88.1 for V4.1-Flash, ahead of every rival with a listed score, according to The Next Web.

The same table shows gaps on hard academic tests, with 36.8 for V4.1-Flash on Humanity's Last Exam against 56.3 for Opus 5, and 20.3 on ProgramBench against 37.0 for Opus 5, according to The Next Web.

Caixin Global noted DeepSeek claims the new model outperforms Anthropic's Opus 5 on some benchmarks, such as Terminal-Bench 3.0, DeepSWE v1.1, CyberGym and Automation-Best, and reportedly outperformed OpenAI's GPT-5.6 Sol on certain benchmarks.

S5 Labs reported DeepSeek lists 90.6 on Terminal-Bench 2.1 and 74.2 on DeepSWE v1.1, but also lists V4.1 Flash at 90.9 on GPQA Diamond against V4 Pro's 92.4, meaning Flash trails Pro on that reasoning evaluation.

The same analysis noted DeepSeek's card reports DeepSWE v1.1 as 74.2 with mini-SWE and 65.6 with Codex, making the harness part of the comparison, and flagged a release-day documentation discrepancy on NL2Repo-Bench at 65.4 in the changelog versus 64.0 in the model card.

The Next Web emphasized these are DeepSeek's figures and said it had not independently verified them, while S5 Labs described them as vendor-reported results where agent scores depend on surrounding software.

Competitive and market context on Sept. 10

The Next Web reported the new rates reverse the direction from August, when DeepSeek launched a Claude Code rival and quadrupled its prices, and cited Bloomberg Intelligence analysts putting the latest cut at as much as 32%.

Bloomberg reported shares in MiniMax and Z.ai fell more than 8% in Hong Kong on Thursday, while Alibaba slid more than 2%.

Caixin Global reported Z.AI in August released and open-sourced GLM-5.3-Flash, priced at about one-tenth of GLM-5.3, claiming stronger performance than GLM-5.2, reflecting a strategy of cheaper efficiency-focused models.

Caixin Global also reported Cambricon Technologies, a Chinese AI chipmaker, achieved Day 0 support for V4.1-Flash, allowing it to run immediately on its hardware upon release.

The Next Web cited Jonathan Kemper at The Decoder, who read the technical report and flagged DeepSeek's description of reward hacking during training, with agents occasionally using newly published vulnerabilities or wiping important system files in test environments, and DeepSeek's admission the model lags the best closed systems at reading complicated images.

Sources referenced