DeepSeek V4.1 Flash cuts KV cache memory fourfold and pushes V4 Pro into retirement

September 10, 2026. DeepSeek formally shipped V4.1 Flash, and it is not just a speed bump. The new model replaces two models at once — V4 Flash and V4 Flash Vision Exp — folds native image understanding into the base architecture, and announces the staged retirement of V4 Pro, whose requests will be routed to Flash starting September 14, 2026 at 04:00 UTC. The message is clear: the product line is collapsing around a single model, and the fight is now over memory and cost.

A multimodal model, bigger yet cheaper to run

V4.1 Flash is a multimodal mixture-of-experts (MoE) model. According to MarkTechPost’s technical breakdown, it runs on 552 billion backbone parameters plus 196 billion “Engram” parameters, with a one-million-token context window. It activates only 8 billion parameters per token during prefill and 16 billion during decode — it is that ratio between total size and activation that keeps per-token cost aggressive.

The structural novelty lies elsewhere: the model cuts KV cache memory fourfold. TechTimes details the four techniques at work — CED split, CSA2, FP4 quantization of the cache, and SWA elimination. In practice, for a long-context agent, the memory needed to track state drops by a factor of four, which directly lowers hosting cost and the latency of long agentic tasks.

Lower prices, and a product line that folds

The notice sent to API users — relayed by TechNode — confirms the pricing strategy: Flash-series prices drop by 11 to 57 percent depending on token type. That is a cost-compression signal more than a frontal price war: DeepSeek is not only cutting the tariff, it is shrinking the resource (KV memory) that made long sessions expensive.

The retirement of V4 Pro is the other half of the message. From September 14, requests addressed to V4 Pro will be routed to Flash. Digital Applied notes that the results favor Flash on many agentic tasks, while Pro keeps the edge on a few. The window until September 14 exists precisely so you can compare on your own workload before the automatic switch.

What it changes for teams

For a team running long-context agents — multi-file research, repository-wide code review, massive log analysis — the fourfold KV cache reduction is the metric that matters most. It changes the economics: what was a GPU-hosting cost that grew with session length becomes more linear, and the one-million-token window becomes genuinely usable in production.

The second point is strategic. DeepSeek is collapsing its line — one multimodal Flash model replaces three entries — and migrating its Pro users onto it. That is a bet on operational efficiency: fewer models to maintain, one to optimize, and a price that reflects the memory savings achieved. For users, the consequence is concrete: you must re-test your loads before the switch, not discover the difference in production.

Four levers to compress KV memory

The fourfold KV cache reduction deserves a closer look, because it is what changes the cost of long agents. TechTimes details the four mechanisms V4.1 Flash combines.

• CED split (Cross-Entropy Distribution split) redistributes the internal representation to cut cache redundancy.
• CSA2 (Cross-layer attention reuse) reuses attention across layers, avoiding storing identical states at multiple depths.
• FP4 quantization of the cache compresses each key and value to four bits, shrinking the footprint relative to FP8 or FP16.
• SWA elimination (sliding-window attention) removes redundant sliding windows that duplicated states.

The combined effect is what makes a one-million-token window economically viable. For an agent reading a whole repository or massive logs, the state-tracking memory stops growing linearly with session length — precisely the bottleneck DeepSeek chose to attack, and a sharper positioning move than a simple price cut.

Open source, and a published comparison

V4.1 Flash ships open source — a point noted by Neowin — and DeepSeek backs it with a 19-benchmark comparison published alongside the announcement. The technical communication is deliberate: the vendor does not just lower prices, it documents where the model wins and where V4 Pro keeps the edge.

Digital Applied sums up the balance: the results favor Flash on many agentic tasks, while Pro still leads on a few specific cases. That honesty is rare in the field, and it is exactly what makes the window until September 14 useful: it is not a migration deadline, but a comparative benchmark window on your own loads.

The subtext: a resource war

Beyond the model itself, V4.1 Flash tells a broader story. AI labs no longer compete only on answer quality, but on the economics of tokens: inference cost, context memory footprint, the price of long sessions. DeepSeek attacks that terrain with FP4 cache quantization and cross-layer attention reuse — optimizations that cut the GPU bill in ways no benchmark shows.

Retiring V4 Pro belongs to the same logic: maintaining a product line is expensive, and concentrating effort on a single multimodal model lets the efficiency gains flow through to price. It is a move other labs are making in their own ways, but DeepSeek makes it explicit here: memory per token has become a first-class metric.

What to watch next

Two follow-ups will determine whether V4.1 Flash becomes the default for agent workloads. The first is real-world behavior on long sessions: a fourfold KV-cache reduction looks strong on paper, but the proof is whether a million-token context degrades gracefully under sustained use rather than in a benchmark. The second is V4 Pro’s actual end-of-life path — how gracefully DeepSeek migrates Pro-dependent workloads and whether the few cases where Pro still leads get a dedicated successor or are simply absorbed into Flash.

For teams, the prudent move is not to wait. Re-run your agentic and long-context evaluations now, record the deltas against your current model, and let the results — not the announcement — decide whether September 14 is a nonevent or a migration.

Verdict

V4.1 Flash marks a shift in where model competition is fought. The comparison is no longer just about benchmarks, but about memory per token and the cost of a long session — that is where DeepSeek is attacking, with a fourfold KV-cache reduction and a line rationalized around a single multimodal model.

If you run long-context agents, the fourfold KV-cache cut and the API price drop make V4.1 Flash an immediate candidate to evaluate — re-benchmark your agentic tasks, where Flash often wins, but keep an eye on the cases where V4 Pro still leads. If you are still on V4 Pro, you have until September 14, 2026 at 04:00 UTC to compare before automatic routing: test your real workloads now, not after the switch. If you are reading the market, note that the race is now over resource (KV memory) and cost, not parameter count alone — and the collapse of the V4 line into a single Flash model is the clearest sign yet of that shift.

References

• TechNode — DeepSeek formally launches V4.1 Flash, routes V4 Pro requests to Flash, September 10, 2026
• MarkTechPost — DeepSeek AI Released DeepSeek-V4.1-Flash with 1M Context, FP4 KV Cache, and Cross-Layer Attention Reuse, September 10, 2026
• TechTimes — DeepSeek V4.1-Flash Cuts Agent Memory Costs Fourfold With New Architecture, September 10, 2026
• Digital Applied — DeepSeek V4.1 Flash: Benchmarks, Prices and Pro Cutoff