DeepSeek V4.1 Flash Is Out — and It's Coming for Its Own Pro Tier

DeepSeek doesn't do slow reveals. Two days ago this was a disposable test build literally named deepseek-v4.1-flash-expires-on-0910. Today — September 10, 2026 — it's a fully released, open-weight model with a technical report on Hugging Face, a self-published benchmark table taking on Claude Opus 5 and GPT-5.6 Sol, and a plan to quietly retire DeepSeek's own Pro tier in favor of it.

Here's what's actually in the release, and what's still just DeepSeek's own claim.

What shipped

DeepSeek-V4.1-Flash is live on Hugging Face under an MIT license, with day-one support in vLLM, SGLang, and Transformers. DeepSeek is calling it the smallest model in a new architecture family — one explicitly designed to scale up into larger models later, the way V3 preceded R1.

The headline architectural change is what DeepSeek calls an asymmetric Causal Encoder–Decoder (CED) design, inspired by the YOCO approach: a 40-layer transformer split into a 20-layer causal encoder and a 20-layer decoder, where the decoder's global KV cache is projected from the encoder's final hidden states rather than computed fresh in every decoder layer. The practical effect is that prompt tokens only pass through the encoder half, roughly halving prefill compute.

On size: DeepSeek's model card puts the backbone at 552B parameters, plus a separate 196B-parameter "Engram" memory module that only activates at two of the forty layers — these aren't meant to be summed into one headline total the way V3's "671B total, 37B activated" was. (Small footnote for the curious: Hugging Face's own repo-size widget lists the checkpoint at 485B params, which doesn't match either figure — an inconsistency sitting on DeepSeek's own page, not something secondary coverage introduced.)

The number that actually matters for your bill is the one everyone agrees on: only 8B parameters activate per token during prefill, 16B during decode. The model also exposes a continuously tunable reasoning effort from 1 to 100, rather than the usual fixed low/medium/high tiers.

It's also DeepSeek's first genuinely natively multimodal model — vision is built into the architecture from pretraining rather than bolted on. The earlier V4-Flash-Vision-Exp, which used a separate vision encoder grafted onto a text-only base, has been retired now that this replaces it.

The number DeepSeek is proudest of: KV cache

A big chunk of the technical report is about shrinking the memory footprint of long-context inference — the thing that actually strains GPUs when you're running agents with huge, repeatedly-reused contexts. DeepSeek's own chart traces the global KV cache size per token across its model generations:

That's roughly a 4x reduction in global KV cache from the previous Flash generation and about 437x smaller than the original DeepSeek-V1 — confirmed directly in DeepSeek's own model card. A second, separate number matters just as much for deployment cost: the model card also states that persistent KV cache (the part that would otherwise sit on SSD) falls to roughly 1/8 of V4-Flash's, because short-lived sliding-window attention state no longer needs to be saved to disk at all — it's cheaply reconstructed by replaying the last 128 tokens on a cache miss instead.

That combination — a new sparse-attention layer-sharing scheme called CSA2, FP4 quantization of the cache itself, and the replay trick above — is what gets both numbers down. Cache-hit tokens are also the cheapest line item in the new pricing below, so this isn't just an engineering flex — it's directly tied to what developers pay.

Benchmarks — DeepSeek's own numbers, with the caveats

DeepSeek published a comparison table running V4.1-Flash against its own V4-Pro and V4-Flash, plus GLM 5.3, Kimi K3, GPT-5.6 Sol, and Claude Opus 5. These are the company's self-reported figures, not independent evaluations, so treat them accordingly — but a few results stand out:

On a separate, harder benchmark generation — Terminal-Bench 4.0 — the gap flips: Opus 5 leads clearly at 51.8 against V4.1-Flash's 31.2, with GPT-5.6 Sol in between at 39.9. And on Humanity's Last Exam, a broad knowledge benchmark, Opus 5's 56.3 comfortably beats V4.1-Flash's 36.8. The honest read: this is a model that's become genuinely competitive with — and sometimes ahead of — flagship Western models on agentic coding and automation tasks, while still trailing them clearly on raw world-knowledge and on the newest, hardest evaluation versions.

Worth knowing before you take the good numbers at face value: this is the same pattern that played out with V4-Pro. DeepSeek's own launch benchmarks put it near the top of the field, but once Artificial Analysis ran it independently, it landed several points behind GPT-5.6 and Kimi K3 on their Intelligence Index — closer to mid-pack than DeepSeek's own chart suggested. That's not an accusation of dishonesty; vendor evals and third-party harnesses routinely disagree by a wide margin for boring methodological reasons like different prompts, scaffolds, and sampling settings.

But it's a reason to treat V4.1-Flash's table as a floor on what independent testers will eventually confirm, not a ceiling — and as of this writing, no independent lab has published a V4.1-Flash score.

Pricing: still the real story

Effective September 10, 2026, DeepSeek's new Flash-tier pricing is:

Peak hours are 01:00–04:00 and 06:00–10:00 UTC on weekdays; everything else — including all of the weekend — bills at the cheaper off-peak rate. For comparison, Claude Opus 5 lists at $5/M input and $25/M output: V4.1-Flash's off-peak output price is over 40x cheaper, and even its peak-hour rate is roughly 20x cheaper.

This is also a reversal, and a fairly quick one. DeepSeek's own API changelog confirms the prior hike took effect at 16:00 UTC on August 16, 2026 — its steepest increase ever — reportedly because ultra-low prices had driven request volume past what its infrastructure could handle, causing instability.

Roughly three and a half weeks later, most of that increase is being walked back: cache-hit input pricing drops by up to 60% in this cut, though output pricing is only trimmed about 11% and remains double the original 2026 launch rate. The likely driver this time: open-weight clones of V4 Flash are now hosted by roughly 28 different providers on OpenRouter alone, many underpricing DeepSeek's own official API.

Pro tier, meet your replacement

DeepSeek is using V4.1-Flash to phase out V4-Pro. Starting September 14, 2026, every request sent to deepseek-v4-pro will be automatically rerouted to V4.1-Flash and billed at Flash's lower rates — DeepSeek says internal and external testers found Flash ahead of Pro on performance, cost, speed, and total completion time combined. The routing stays in place until a V4.1-Pro model eventually ships, with no date announced for that yet.

If you're building on top of V4-Pro today, this is worth planning around even though it requires no code changes on your end: the model actually answering your requests is about to change underneath you.

The bigger picture

None of this is happening in isolation, and the pieces connect more tightly than they might first appear. On September 9 — the day before this model shipped — Reuters reported DeepSeek has hired CITIC Securities to prepare for a potential IPO on Shanghai's STAR Market, aiming to begin the process this year; the same reporting notes DeepSeek is separately closing a funding round at a roughly 500-billion-yuan ($75 billion) valuation.

Two days before that, a DeepSeek technical lead announced the company is opening about 150 new roles — specifically backend and server-side engineers, not model researchers — to rebuild infrastructure straining under rapid growth in users and compute-heavy agent workloads.

Put together, that's a company whose backend is visibly under strain from demand (hence the August price hike), that's throwing headcount at fixing it, that just shipped a model engineered specifically to cut inference infrastructure costs (that KV-cache work isn't incidental), and that's doing all of it while lining up both private funding and a public listing. A cheap, efficient, open-weight model that quietly retires its own pricier Pro tier fits that picture well — it's as much an infrastructure-cost story as a product launch.

Bottom line

V4.1-Flash isn't uniformly "better than everything," and DeepSeek's own numbers say as much once you look past the headline wins — Opus 5 and GPT-5.6 Sol still lead clearly on raw knowledge and on the newest, hardest agentic benchmarks, and the V4-Pro precedent says even the wins deserve a discount until someone outside DeepSeek reruns them.

But as a cost story, it doesn't need the benchmark table to make its case: a model landing anywhere near $5–25/M-token flagships, priced at cents, open-weighted under MIT, shipped the same week its own maker was quietly lining up an IPO and firefighting its own server load. Judge the model on the numbers once independent labs weigh in. Judge the moment on what a company does when it's simultaneously scaling, straining, and trying to look investor-ready — and by that measure, this was a confident week.

The architecture and KV-cache figures above are confirmed directly against DeepSeek's own model card and technical report on Hugging Face. The benchmark table and pricing are DeepSeek's self-reported figures, corroborated across multiple outlets that reviewed the same release materials — independent, third-party benchmark runs had not yet published at the time of writing.