DeepSeek V4.1 Flash launch lands under a version number that downplays what the company actually shipped
The DeepSeek V4.1 Flash launch arrived on September 10, 2026, with MIT-licensed weights, a one-million-token context window, and a wholly new Causal Encoder-Decoder architecture sitting on top of a 552-billion-parameter mixture-of-experts backbone. DeepSeek’s own model card files the model under what the company is calling a “new architecture family,” and the size of the change is hard to reconcile with the modest “.1” version number the company chose.
Version bumps, in DeepSeek’s own nomenclature, normally extend an existing family. The V4.1 release does something else: it retires a 1.6-trillion-parameter flagship in favor of a Flash-tier model and, in the process, replaces the architecture that family was built on. DeepSeek V4.1 Pro, the flagship that the new numbering implies should exist, is not listed in the company’s materials.
Why a patch-number release reads like a generational one
One independent tracker framed the “V4 Flash to V4.1 Flash” transition as a patch step and recommended budgeting accordingly. The numbers underneath that recommendation disagree. The architectural choices DeepSeek published are not the choices a tuning pass would make, and the decision to retire the prior flagship rather than ship V4.1 alongside it is a louder signal still.
Four components of the new model stand out in the model card. The release uses FP4 main KV caching at 890 bytes per token, which DeepSeek’s own writeup puts at roughly a quarter of the prior DeepSeek V4 Flash baseline. Chinese coverage has pushed that comparison further, framing the compression as a 437× reduction against the first-generation V4 model, but that larger figure is vendor arithmetic on a different baseline and should be treated as a claim rather than a measurement.
The second piece is Compressed Sparse Attention 2, in which each attention layer runs in one of three static modes — Full, Reindex, or Reuse — with KV and indexer state shared across layers and a hierarchical sparse indexer bounding cost independently of context length. The card also documents an inference folder and a separate encoding module in the open-source repository, with DeepSeek explicitly inviting the community to build out serving. That phrasing is the standard signal that day-zero support in the major runtimes is days away rather than weeks.
The previous experimental endpoint, suffixed “-expires-on-0910,” ended on schedule. The real release landed the same day across DeepSeek’s web app, mobile app, and first-party API, with weights and a technical report pushed to Hugging Face under deepseek-ai/DeepSeek-V4.1-Flash.
Benchmarks, pricing, and a quiet model-name rewrite
On DeepSeek’s own benchmark sheet, V4.1 Flash posts 90.9 on GPQA Diamond, a Codeforces rating of 3471, 65.6 on MathArena Apex, 90.6 on Terminal-Bench 2.1, 74.2 on DeepSWE v1.1, and 63.9 on HLE with tools. The HLE figure carries a footnote that limits the comparison baseline of 36.8 to the pure-text subset of that benchmark.
The same vendor reporting has V4.1 Flash winning 13 of 16 comparisons against Kimi K3 and 11 of 13 against GLM-5.3, while still trailing GPT-5.6 Sol and Claude Opus 5 on the newer Terminal-Bench 3.0 and 4.0 tasks and on ProgramBench. The shape is coherent — strongest at the coding, terminal, and agent-automation work the new architecture is optimized for, weaker at frontier-reasoning tasks — and it is the shape a generational step would plausibly take.
The API pricing on the new model is set in two tiers, with peak windows running 01:00–04:00 and 06:00–10:00 UTC on weekdays. Cache-hit input runs $0.003 per million tokens off-peak and $0.006 at peak, cache-miss input $0.15 and $0.30, and output $0.60 and $1.20 — peak priced at exactly double off-peak across the board.
Alongside the model, DeepSeek quietly rewrote how legacy aliases resolve. The recommended model name is now deepseek-flash, while the prior deepseek-v4-flash and deepseek-v4-flash-vision-exp strings are still accepted but no longer point at the models they used to; both are retired, and any request that names them is served by DeepSeek V4.1 Flash and billed at the Flash rate. The same routing change applies to deepseek-v4-pro, which now also resolves to V4.1 Flash and stays there until DeepSeek V4.1 Pro ships.
The economics of that rerouting are sharp. DeepSeek V4 Pro’s list pricing is $0.022 and $0.044 cache-hit input, $0.66 and $1.32 cache-miss input, and $1.98 and $3.96 output. Sending a Pro-named request to V4.1 Flash instead drops the output cost by roughly 70% while, by DeepSeek’s own accounting, raising the capability of the model answering. Integrations that hard-code the Pro model name will not break; they will receive a different model at roughly a third of the prior output price, with no code change and no notice beyond a changelog entry. Integrations that route by capability tier rather than by name have until September 14 to re-test their routing logic.
For inference providers, the practical question is whether the architectural shift and the pricing reset arrive on the same day in their catalogues. CoinCustard routes the rest of the DeepSeek line — DeepSeek V4 Flash and DeepSeek V4 Pro, alongside more than 200 other models — under the vendor’s published list price with zero markup, so any rate change DeepSeek makes is live on our side the same day rather than after a repricing cycle. When V4.1 Flash does land in the catalogue, the deepseek-flash alias is the one customers should target, and the deepseek-v4-pro string is the one they should stop targeting — making the DeepSeek V4.1 Flash launch as much a routing event as a model release.
Source: https://www.coincustard.com/news/deepseek-v4-1-flash-launch