DeepSeek V4.1 Flash: Release Date, the Test Model That Expires September 10, and What DeepSeek Has Published

§ 01What we actually know#

§ 02What the notice said, and who carried it#

DeepSeek did not publish the notice on a public page, so the primary record is the press that quoted it. IT之家 posted at 16:05 Beijing time, citing readers who saw the message in DeepSeek’s official group: an intermediate version of DeepSeek V4.1 Flash is in internal testing, it uses a new model structure with native multimodal support, it is stronger, faster and cheaper, testers keep base_url unchanged and set the model to deepseek-v4.1-flash-expires-on-0910, billing matches deepseek-v4-flash, and each account is limited to 20 concurrent requests. IT之家 also reported the feedback form’s question about replacing V4 Pro. Phoenix Tech carried the same points at 16:17. Machine Heart, at 18:00, put the group message at about 3 pm and added that “internal test” was generous: any API user could add the model name and call it.

Three things about the notice are worth reading carefully. First, “intermediate version” is DeepSeek’s own framing. This is not the release model; it is a checkpoint offered for feedback, and the expiry in the id enforces that. Second, “new model structure” is a stronger claim than DeepSeek made for V4-Flash-0731, whose model card describes it as the official release of V4-Flash with enhanced agentic capabilities and the same structure as the DSpark variant. A structural change between 4 and 4.1 is a bigger step than the version number suggests. Third, “native multimodal” is a distinction from the current approach. V4-Flash-Vision-Exp added a vision encoder and aligner to the text model and kept training. Native means the images are part of the base design. If true, it explains why DeepSeek would call it a new structure.

§ 03The rumor chain, dated#

The chain is short, because DeepSeek does not leak the way Western labs do. What it does instead is ship with the date in the file name.

The expiry is the only date, and it is not a launch date. expires-on-0910 tells testers when the checkpoint disappears. It does not tell anyone when a release model arrives, and DeepSeek’s dated suffixes have always described the release itself (0731, 0813), never a countdown. Reading September 10 as a release date is the mistake this page exists to prevent.

The speed numbers are one person’s tests. Machine Heart credits the 400 tokens per second figure to unnamed community members and the 3.9 to 6 times speedups to X user @NFT_Chen, comparing the test model against V4-Flash-Vision-Exp on a 49K-token retrieval task (5.2 times) and SVG code generation (6 times). Those are plausible for a model DeepSeek itself calls faster, and they are self-tests of an intermediate checkpoint over a two-day window. We list them as reported. The Gemini 3.8 Flash comparison Machine Heart draws is the comparison DeepSeek’s launch table will eventually make or avoid; until then it is a tweet.

The V4 Pro question is the real signal. DeepSeek asked its testers whether this Flash-class model can fully replace V4 Pro online. V4-Pro is the 1.6T total, 49B active flagship; V4-Flash is 284B total, 13B active. If DeepSeek is seriously weighing a Flash replacing Pro in its own service, the new structure is doing a great deal of work, and it would fit a company whose April announcement led with “cost-effective 1M context” rather than raw scale.

§ 04DeepSeek’s cadence#

The reason a September release is plausible has nothing to do with the notice. It is the calendar.

Dates come from the DeepSeek API news index and change log; V4-Flash-0731 is dated by its July 31 change log entry, and V3-0324 by its news page rather than the March 24 in its name. Two readings of the table. The long gaps are between generations: 144 days from V3.2 to V4 Preview, 98 from the preview to the first final V4 model. The short gaps are inside a generation: 13 days from V4-Flash-0731 to V4-Pro, 8 from V4-Pro to Vision-Exp. A 4.1 that is a structural change sits somewhere between those two rhythms. The September 8 test is 18 days after Vision-Exp; if V4.1 Flash ships in September, it will be the fastest structural update DeepSeek has made, and the expiring test id is the first time DeepSeek has previewed a model this way rather than shipping it.

§ 05What 4.1 would land on top of#

The V4 lineup is well documented, which is what makes the 4.1 silence stand out. DeepSeek’s April 24 announcement gave the architecture: V4-Pro at 1.6T total and 49B active parameters, V4-Flash at 284B total and 13B active, both with a 1M context as the default across all official DeepSeek services, built on token-wise compression plus DeepSeek Sparse Attention. The July 31 V4-Flash-0731 card calls itself the official release of V4-Flash and reports it beating the V4-Pro preview on agent benchmarks despite the far smaller active count (Terminal Bench 2.1 at 82.7, DeepSWE at 54.4, Toolathlon-Verified at 70.3 on DeepSeek’s own table). The card’s weight total reads about 304B because the release ships with a speculative decoding module attached; the announced base is 284B. The August 13 V4-Pro GA post added reasoning effort levels for both models, native OpenAI Responses API support, and the pricing change: peak and off-peak rates, off-peak 50 percent lower, effective August 16 at 16:00 UTC. Current figures are on the pricing page; the test id is billed at whatever deepseek-v4-flash costs at the hour you call it.

The agent tooling is moving too. DeepSeek Harness 0.1.1 shipped with Vision-Exp on August 21, and the GitHub repository tagged 0.1.3 pre-release builds every day from September 1 to 7. A new model with a new structure arriving alongside a harness release would match how V4-Pro landed. Our record of the wider open-weight wave this lands in is in the Qwen3.8-Flash-Next and GLM-5.3-Flash posts.

§ 06What we are watching for#

We run production work on three model families and read the open-weight labs closely, so this list is an operator’s.

• An entry on the API news page. Every DeepSeek release since V3 has one. The moment a 4.1 page appears at api-docs.deepseek.com/news, this stops being a rumor page.
• A permanent model id. deepseek-v4.1-flash without an expiry, or a dated release suffix like the ones on 0731 and 0813, on the models and pricing page.
• Weights and a card on Hugging Face. V4-Flash-Vision-Exp’s weights followed its API release by ten days. A card would give the first parameter count and confirm or retire the “new structure” line.
• September 10. If the test id expires and nothing replaces it, that is a scorecard row. If a second test id appears with a later expiry, that is a pattern.
• A price. “Lower cost” is in the notice. DeepSeek’s V4 pricing change was announced eight days before it took effect; a 4.1 price would likely arrive the same way.
• The V4 Pro decision. Whether DeepSeek’s release post says anything about the online service’s default model, and whether deepseek-v4-pro keeps pointing at 0813.
• Multimodal spec. Whether “native” means the release model accepts images without a separate vision-exp id, which would retire the August 21 experiment.

§ 07The tracker#

This page updates when facts change, not when posts get louder. As of September 8, 2026: DeepSeek V4.1 Flash exists as a test model id with a two-day life, described in a group notice, tested by strangers and announced nowhere. No release date, no spec, no benchmark and no price exist in any DeepSeek document. The likeliest picture, on DeepSeek’s own rhythm, is a release in September with the date in its name. When DeepSeek moves, the confirmed facts and a graded scorecard of every claim above land here the same day.

§ 08Update log#

This is a living page; when the story moves, the update lands here.

As of September 8, 2026: page opened. Test id deepseek-v4.1-flash-expires-on-0910 live per DeepSeek’s community notice; no official announcement to log.