Analysis · 2026-09-15 · 9 min read

DeepSeek-V4.1-Flash: Specs, Speed and Token Cost of the New Multimodal MoE

DeepSeek's V4.1-Flash is a native multimodal Mixture-of-Experts model with a 552B backbone, 1M-token context and 300-400+ tokens/sec throughput. Here's the architecture, the speed numbers, and what it actually costs per million tokens.

TL;DR

DeepSeek-V4.1-Flash is a fast, low-cost, native multimodal Mixture-of-Experts model released by DeepSeek in September 2026.
The design is asymmetric: a 40-layer causal encoder-decoder with a 552B-parameter backbone plus 196B parameters of Engram conditional memory, yet it activates only 8B parameters per token during prefill and 16B during decode.
1,000,000-token context, Compressed Sparse Attention 2 and an FP4 main KV cache cut the global KV cache footprint dramatically — which is why a model this large can serve this cheaply.
Real-world throughput lands at 300 to 400+ tokens per second, with a continuously adjustable reasoning effort dial from 1 to 100.
On our pricing table, DeepSeek V4.1 runs $0.42 / $1.68 per million input/output tokens — roughly 4% of a frontier US flagship's output price.

!DeepSeek logo

!DeepSeek on OpenRouter

!NVIDIA

1. What DeepSeek Actually Shipped

V4.1-Flash is not a minor point release on top of V4 Flash. It is a re-architected model whose entire premise is higher intelligence per gigabyte of serving memory — and memory, not FLOPs, is what sets the floor on inference prices in 2026.

The headline choice is asymmetry. Instead of a uniform stack of identical transformer blocks, V4.1-Flash splits capacity between a heavyweight backbone and a separate conditional-memory system that is only consulted when the router decides it is worth it. The result is a model with frontier-scale knowledge that behaves, at runtime, like a small one.

2. Key Specifications

Two numbers deserve a second look.

8B prefill / 16B decode active parameters. Total capacity is 748B across backbone and Engram modules, but only single-digit-to-low-teens billions are touched per token. That ratio — roughly 2% activation — is the whole cost story. You pay for the intelligence of a 748B model at the serving economics of a mid-size dense one.

Asymmetric prefill and decode. Most MoE models activate the same expert budget in both phases. V4.1-Flash deliberately spends *less* during prefill (where you are simply ingesting a possibly million-token prompt) and *more* during decode (where each token is a real decision). For long-context, short-answer workloads — RAG, document QA, codebase analysis — that asymmetry is a direct discount.

3. Why the Memory Footprint Matters More Than the Parameter Count

At million-token context lengths, the KV cache, not the weights, dominates GPU memory. A conventional FP16 KV cache at 1M tokens is a serving-economics disaster: it evicts other requests, kills batch size, and drives the per-token price up.

V4.1-Flash attacks this on two fronts:

Compressed Sparse Attention 2 reduces how much attention state must be retained at all, rather than compressing it after the fact.
An FP4 main KV cache stores what remains at four bits, cutting the resident footprint against an FP16 baseline by roughly 4x.

Combined, these shrink the *global* KV cache footprint enough that a provider can hold far more concurrent long-context sessions on the same hardware. High batch size is exactly what makes a low posted price sustainable rather than a loss-leader.

4. Performance and Speed

DeepSeek reports 300 to over 400 tokens per second in real-world benchmarks — not a lab-only figure measured at batch size 1 with an empty cache. For interactive agents, that is the difference between a UI that streams faster than a person reads and one that visibly stalls.

Where V4.1-Flash improves most over the previous generation:

Coding. Materially stronger on multi-file edits and long-horizon refactors, which is where the 1M context and the decode-heavy activation budget compound.
Spatial reasoning. Native multimodality, rather than a bolted-on vision adapter, shows up here — the model reasons about layout, geometry and relative position rather than captioning.
3D environment generation. Scene graphs, procedural layouts and consistent multi-view descriptions are a specific target of this release.

The Reasoning Effort Dial

The 1-100 continuously adjustable reasoning effort setting is the most practically important feature for anyone watching a budget. Competing models expose a coarse low/medium/high switch; a continuous dial lets you tune spend per route rather than per product.

Reasoning effort is a *token multiplier*, not a price multiplier — the posted per-token rate never changes. Turning the dial from 20 to 90 can still multiply your bill several times over, because the model simply emits more reasoning tokens. Set it per route and log the distribution.

5. Token Cost

Current published rates for DeepSeek V4.1 on our pricing table:

What that means in practice, for a typical agent turn with a 30,000-token prompt and 1,500 output tokens:

Uncached: (30,000 / 1M x $0.42) + (1,500 / 1M x $1.68) = $0.0151 per turn
With a warm cache on the 28,000-token static prefix: roughly $0.0045 per turn — about a 70% saving

At one million such turns a month, that is the difference between ~$15,100 and ~$4,500. The caching discipline described in Prompt Caching Explained is worth more than most model swaps.

6. Where V4.1-Flash Fits

Use it for:

High-volume production traffic where quality is good enough and price per token decides the architecture.
Long-context workloads — full codebases, contract sets, multi-hour transcripts — where the FP4 cache and CSA2 keep 1M-token prompts economically sane.
Multimodal pipelines that need image and layout understanding without stitching together a separate vision model.
Latency-sensitive agents that need 300-400+ tokens/sec streaming.

Think twice for:

Frontier-hard reasoning where a top-tier flagship still wins the last few points of accuracy and the task value dwarfs the token cost.
Strict data-residency setups that require self-hosting — feasible with open weights, but 748B total parameters is a real hardware commitment even at 2% activation.

7. The Bottom Line

DeepSeek-V4.1-Flash is the clearest statement yet that the interesting frontier in 2026 is memory efficiency, not parameter count. A 552B backbone plus 196B of conditional memory that activates 8-16B per token, holds a million-token context in an FP4 cache, and streams past 400 tokens per second is a fundamentally different cost curve than a dense flagship.

For most production workloads the question is no longer "can a cheap model do this?" but "which of my routes genuinely need the expensive one?" Price the routes, set the reasoning dial per route, cache the static prefix, and V4.1-Flash will handle a surprising share of them.