DeepSeek V4.1 Flash: 890 Bytes KV Cache per Token

Overview

On September 10, 2026, DeepSeek released V4.1-Flash, the third model in the V4 family — with a clear message: it aims to simultaneously surpass the previous Flash toolkit in both quality and cost. The numbers show exactly that. With 552 billion backbone parameters and 196 billion engram parameters, the model activates only 8 billion per token during prefill and 16 billion during decoding — featuring a context window of one million tokens and a maximum output of 384,000 tokens. The global KV cache drops to 890 bytes per token. The tech report is fittingly titled "Pushing the Limits of KV Cache Compression."

The weights are available under the MIT license on Hugging Face, and it can be served via vLLM, SGLang, and Transformers. API pricing remains at the V4-Flash level, utilizing a peak/off-peak pricing model.

What the Model Is

V4.1-Flash is a multimodal Mixture-of-Experts (MoE) model. It processes text and images natively — a DeepSeek-ViT with 2D-RoPE and 3x3 pixel unshuffle provides visual embeddings that are jointly trained with text from the very first pretraining step. Per MoE layer, 1 shared and 384 routed experts are at work, with 6 active per token. Reasoning is seamlessly adjustable (integer 1–100): more "thinking effort" yields measurably better results but costs about 2.5 times as many output tokens.

The Architecture: Three Levers Against the Cache

The bottleneck in modern agent systems is rarely model size, but rather the KV cache: every layer of a classical Transformer maintains its own key-value pairs, and with long contexts, this buffer balloons into the gigabyte range. DeepSeek tackles this on three fronts.

First: Causal Encoder-Decoder (CED). The 40-layer backbone is split into 20 encoder and 20 decoder layers. The decoder does not compute its own global KV cache; instead, it projects it from the final encoder hidden states. Prompt tokens essentially stop in the encoder — the prefill compute load drops to about half, with only 8 billion parameters active there. The inspiration here is YOCO.

Second: Compressed Sparse Attention 2 (CSA2). Each CSA2 layer is statically assigned one of three modes: Full (own main KV, own indexer, fresh top-512 indices), Reindex (borrowed main KV, own re-evaluation), or Reuse (main KV and indices completely adopted). The 18 encoder CSA2 layers compress at a ratio of 2 in groups of six; the 20 decoder layers at a ratio of 1 in groups of four. A hierarchical sparse indexer limits the candidate pool to 16,384 positions — meaning the indexing effort no longer scales linearly with context length. In terms of FLOPs per decode token, the jump from a 4K to 1M context costs only a quarter more.

Third: FP4 and Bounded Replay. The main KV cache is quantized to E2M1 (one E4M3 scale per 16 channels), inspired by NVFP4 — which halves it once again. The sliding-window states (128-token window) no longer land on the SSD but in the host RAM pool with a minute-level TTL; in case of a cache miss, the "SWA Bounded Replay" simply recompiles the last 128 tokens. The result: the persistent cache footprint is about one-eighth of V4-Flash, globally 890 bytes per token — roughly 1/4 of V4-Flash and 437 times smaller than V1. The tech report details the progression: 389,120 → 48,068 → 3,514 → 890 bytes.

The rest of the package consists of proven building blocks: single-pass mHC with shifted input-mixing coefficients and a fused mega-mHC kernel, the Engram module for token-based lookup in layers 1 and 14, and DSpark for speculative decoding with confidence-scheduled verification.

Cross-Section of Performance

In the base comparison, V4.1-Flash holds its own against the much larger V4-Pro — with one-third of the total parameters and one-quarter of the activated ones: MMLU-Pro 74.1 (Pro: 73.5), BigCodeBench 60.6 (59.2), GSM8K 93.0. Natively multimodal: DocVQA 95.6, CVBench 77.9, RefCOCO 86.0, MMMU-Pro 56.5.

At the maximum reasoning level on the agent benchmarks (Source: Model Card):

Additionally: Codeforces rating 3471 (V4-Pro: 3348), MathArena Apex 65.6 (same as Kimi K3). On ProgramBench and when reading complex images, the report notes a measurable gap remains compared to the largest closed systems.

The post-training story is remarkably sober: no new algorithms, just SFT → RL → on-policy distillation using over 40 teacher models and scaled, synthesized agent environments in the Claude Code, Codex, OpenCode, Pi, mini-SWE, and DeepSeek Harness scaffolds. DeepSeek attributes the gains to data and task scaling — not algorithm updates.

Pricing and Availability

The API serves V4.1-Flash at the same rates as V4-Flash, utilizing peak and off-peak windows; off-peak costs half. Per million tokens: $0.22 input (cache miss), $0.007 for a cache hit, $0.66 output — all off-peak, with peak rates being exactly double. Automatic prefix caching is enabled by default. The new rates have been active since 04:00 UTC on September 10, 2026. Weights: MIT License, Hugging Face. Official partners like WorkBuddy (including CodeBuddy) and OpenCode support the model natively at launch.

Practical Implications

For long documents and knowledge graphs: The 437-fold shrinkage in cache per token is the real breakthrough. Those working with 1M contexts and frequent tool calls can scale more sessions per GPU instead of having to upgrade the GPU itself.

For agent pipelines: The asymmetric activation (8B prefill / 16B decode) perfectly matches input-heavy patterns — large amounts of source code, short, structured outputs. The continuously adjustable reasoning effort axis allows users to hit the cost-benefit sweet spot at every step: low level for numbering, high for cross-checking, medium for packaging.

For operations: In vLLM, the mixed FP4/FP8 checkpoint applies; use --tokenizer-mode deepseek_v4, --reasoning-parser deepseek_v4, and for "think" modes, --max-model-len must be at least 393,216. Two recommendations from the Nvidia forum: 32K context remains the stable default; 64K should only be used if truly needed — the KV utilization and real-world concurrency show a clear drop-off point.

For industry context: In this generation, open-weights models are no longer just "good for the price" but are top performers in four out of five hardcore agent benchmarks — under an MIT license. Gaps remain in fine-grained image detail questions and scientifically deep reasoning chains, where V4-Pro, with 49B activated parameters, remains the safest bet.

Key Takeaways

1. CED moves the prompt to the encoder: 8B active parameters in prefill, 16B in decode, the decoder KV comes as a projection.
2. CSA2 + FP4 + Bounded Replay squeeze the global KV cache down to 890 bytes per token: 1/4 of V4-Flash, 1/8 of the persistent portion, 437× smaller than V1.
3. Benchmarks: Terminal-Bench 2.1 90.6 · DeepSWE v1.1 74.2 · AutomationBench 54.8 · CyberGym 88.1 — with MIT weights.
4. Operations: vLLM/SGLang/Transformers, automatic prefix caching, 32K context as a stable default.

FAQ

Is V4.1-Flash a complete replacement for V4-Pro? In agentic coding benchmarks, yes, and significantly so regarding price and speed. For scientifically dense reasoning chains and complex image analysis, the model card still lists a lead for the larger models.

Why is the KV cache the bottleneck? Because every layer of a classical Transformer maintains its own key-value pairs — this buffer grows with context length and steps, tying up HBM, SSD, and bandwidth.

How do I adjust the reasoning? Via reasoning_effort as an integer from 1–100; the maximum level measurably improves results but roughly doubles the output tokens.

What are the concrete figures for the cache advantage? The global footprint drops from 3,514 bytes per token (V4-Flash) to 890 — a factor of 4, with a 437-fold difference compared to V1.

Which serving stack should I use? vLLM or SGLang with deepseek_v4 tokenizer and reasoning parser; FP4/FP8 mixed checkpoint; think modes require max-model-len ≥ 393,216.