DeepSeek's own official account, 深度求索 — the company's Chinese name, run as its community/support channel — confirmed today what several independent sightings had already described, relayed in a screenshot posted by the account @MaxForAI: "DeepSeek V4.1 Flash's interim version has entered internal testing, feel free to try it," followed by the identical technical details third-party accounts had already reported — new architecture, native multimodal support, "more capable, faster, and lower cost" — and a request for feedback via a linked survey form. Chinese tech outlets covering the same test report, secondhand, that the feedback DeepSeek is specifically soliciting centers on one question: whether this Flash-tier model can substitute for V4-Pro entirely — meaning the goal isn't just a cheaper model, but Pro-level capability at Flash's cost and speed. That's not confirmed by DeepSeek's own post directly, but it would explain why a company iterates on its budget tier with a "new architecture" claim rather than its flagship. The model is reachable by keeping the same API base URL and setting the model name to deepseek-v4.1-flash-expires-on-0910, priced the same as the current deepseek-v4-flash tier, capped at 20 concurrent requests per account.
Third-party sightings and DeepSeek's own account now agree, detail for detail
Before DeepSeek's own confirmation surfaced, at least three independent accounts had already described the identical technical specifics — the exact model string, the "expires-on-0910" suffix, and the instruction to keep the existing base URL and just change the model name — which was already a meaningfully stronger signal than a single unverified post. DeepSeek's own community-account post matches those third-party reports point for point: same model string, same pricing note, same 20-concurrent-request cap, plus the "new architecture" and "native multimodal support" language. That agreement is reassuring on the facts, but it doesn't change what's still missing: no pricing has actually changed from the current deepseek-v4-flash tier, so any "cheaper to run" claim has to come from doing a given task in fewer tokens or fewer calls, not from a headline price cut — a real efficiency claim, but a different one than "cheaper" suggests on its own.
An official beta, not an official launch
A model ID with a hard-coded expiration date, distributed through a community account with a feedback survey attached, is still not how DeepSeek normally announces a product. Compare this to DeepSeek's own V4-Pro launch last month, which arrived via a post from DeepSeek's main account, came with a full benchmark table against named rivals, and gave the new checkpoint a permanent model name that existing integrations didn't need to change to pick up. Today's confirmation has none of that: no benchmark table, no permanent name, and an explicit two-day test window. DeepSeek is officially confirming a beta, not officially launching a model — a real distinction, not a technicality, since everything about how this is being distributed (temporary ID, capped concurrency, a survey asking testers whether it can replace the Pro tier) says DeepSeek itself doesn't yet consider this checkpoint finished.
If "native multimodal" holds up, that's the actual story
DeepSeek's V4 line, as this blog covered at its Pro launch, has competed primarily on agentic coding and reasoning benchmarks against Kimi K3, Claude Fable 5, and Claude Opus 4.8 — a text-and-code competitive set. Native multimodal support would be a genuine architectural shift for that line, not an incremental update, and it would put DeepSeek in more direct competition with the vision-capable frontier models this blog has tracked all month. That's exactly why it's worth treating as an interesting claim to verify rather than a confirmed capability: "native multimodal support" attached to a two-day beta test with a feedback survey, and zero benchmark numbers attached from DeepSeek or anyone else, is not yet the same thing as a shipped, evaluated capability.
What to expect next
- Watch whether the endpoint actually disappears on September 10, as named. If it does, that confirms this stayed a bounded test; if DeepSeek extends it or converts it into a permanent model name, that's the tell that testing went well.
- Watch for the eventual official launch post and benchmark table. DeepSeek's V4-Pro launch set the template — a main-account post with a full comparison table against named rivals; that's the actual test of whether V4.1 Flash's claims hold up, not today's beta confirmation.
- Watch for independent benchmark results before crediting "faster," "more capable," or "cheaper." All three claims currently rest on DeepSeek's own beta-announcement language, not on any measured comparison from DeepSeek or an outside evaluator.
- Watch whether the survey question reported secondhand — can Flash replace Pro entirely — gets answered explicitly in whatever DeepSeek says when the test period ends, since that's the more consequential claim than any individual benchmark score.
Update — September 10: the official launch, with the benchmark table this post said to watch for
Two days after the beta confirmation above — right at the edge of the "expires-on-0910" window the temporary model ID itself named — DeepSeek's main account, @deepseek_ai, posted the official V4.1 Flash launch: "the smallest model in our new architecture family, with native visual understanding," alongside a benchmark chart against Kimi K3, GLM-5.3, Claude Opus 5, and GPT-5.6 Sol across four tasks (Terminal-Bench 3.0, DeepSWE v1.1, CyberGym, and Automation-Bench). That's exactly the format this post said would be the real test — a main-account post with named rivals — rather than the community-account beta confirmation covered above, and DeepSeek made the transition cleanly: the beta model ID's own self-imposed expiration date lines up almost exactly with the official launch date.
On DeepSeek's own chart, V4.1 Flash leads three of the four bars — CyberGym (88.1, ahead of every named rival), Automation-Bench (54.8), and DeepSWE v1.1 (74.2, edging Claude Opus 5's 74.0 by a fraction of a point) — but is clearly behind on the fourth: Terminal-Bench 3.0, where it scores 30.0 against Opus 5's 43.3 and GPT-5.6 Sol's 34.4, though still ahead of Kimi K3 (17.7) and GLM-5.3 (28.3). DeepSeek didn't drop or reframe that bar even though it's the one place its own model trails two of four named competitors by a wide margin — worth crediting as the kind of complete disclosure this post noted was absent from the beta announcement two days earlier.
What the chart doesn't touch is the headline architectural claim from the same launch post: "native visual understanding." All four benchmarks shown are text-and-agentic tasks — terminal use, software engineering, cyber exercises, and automation — not vision tasks, so the multimodal claim this post flagged as the actual story if it held up remains exactly where it was: asserted by DeepSeek, untested by any benchmark DeepSeek has shown so far. The launch announcement is the first post of a six-part thread; only the opening post and its chart were available at the time of writing, so pricing, access details, and the permanent model name may appear in the later parts of the thread not covered here.
A fuller benchmark table posted later in the same thread makes the more consequential comparison explicit: V4.1 Flash against DeepSeek's own previous-generation flagship, V4-Pro-0813, not just against outside rivals. Across roughly a dozen shared benchmarks — Terminal-Bench 2.1/3.0/4.0, DeepSWE v1.1, ProgramBench, NL2Repo-Bench, CyberGym, SEC-Bench Pro, ExploitGym, HLE with tools, Automation-Bench, Agents' Last Exam, and Codeforces rating — V4.1 Flash beats or ties the model DeepSeek previously sold as its flagship, in some cases by a wide margin (ExploitGym: 15.3 vs. 5.4; Terminal-Bench 3.0: 30.0 vs. 11.8). Pro-0813 still leads clearly on two general-knowledge benchmarks: GPQA Diamond (92.4 vs. 90.9) and HLE (42.7 vs. 36.8, or 39.1 on V4.1 Flash's text-only subset). That's a direct answer to the question this post flagged as more consequential than any single benchmark score two days ago — whether Flash could replace Pro entirely: DeepSeek's own numbers now show its smallest new-architecture model can, on most of the tasks it chose to publish, replace the tier above it, with the clear exception of the broadest knowledge-recall evaluations.
That same fuller table also revises something said just paragraphs above: that none of the benchmarks DeepSeek had shown tested the "native visual understanding" claim. The complete table includes three benchmarks with vision-flavored names — Chartography, BabyVision, and ZeroBench-main, all run with tools — where V4.1 Flash posts real, non-zero scores (78.9, 89.6, and 49.0) that beat Kimi K3 on all three but trail Claude Opus 5 on all three (84.0, 94.1, 52.0) and roughly match GPT-5.6 Sol. GLM 5.3 has no listed score on any of the three. The multimodal claim isn't untested after all — it's tested, and the results describe a real but not class-leading performer on whatever those three evaluations measure, a more precise picture than "asserted, untested" a few hours earlier.
DeepSeek also backed the "faster inference, higher throughput" claim from the launch post with a concrete mechanism: a chart showing global KV cache size per token falling from 389,120 bytes in the original DeepSeek-V1 to 890 bytes in V4.1 Flash — an 8.1x cut to V3.2, a further 13.7x cut to V4-Flash, and a further 3.9x cut to V4.1 Flash, each drop directly reducing the memory bandwidth inference requires. And where the beta covered above matched the flat rate of the existing V4-Flash tier, the official API pricing, effective 04:00 UTC on September 10, introduces a structural change: distinct peak and off-peak rates, exactly double during peak hours (01:00–04:00 and 06:00–10:00 UTC on weekdays; everything else, including weekends and holidays, is off-peak) across all three price components — $0.003 versus $0.006 per million input tokens on a cache hit, $0.15 versus $0.3 on a cache miss, and $0.6 versus $1.2 for output. That's not the flat single rate the beta announcement described two days ago; it's a new, time-of-day-dependent structure specific to the official launch.