DeepSeek V4.1-Flash: The Asymmetric Architecture That Killed Its Own Flagship
On September 10, 2026, DeepSeek did something unusual in the AI industry: it released a model so good that it immediately announced the retirement of its own flagship. DeepSeek-V4.1-Flash, a 552B-parameter Mixture-of-Experts model with a brand-new Causal Encoder-Decoder architecture, didn’t just match the V4 Pro — it surpassed it on performance, cost, speed, and total runtime. By September 14, every request to deepseek-v4-pro would be silently routed to V4.1-Flash at Flash pricing.
This is not a routine version bump. It is a structural redesign that rethinks how compute is allocated during inference, and it has implications for every developer building agents on the DeepThink platform.
The Core Innovation: Asymmetric Activation
The single most important design decision in V4.1-Flash is input-output asymmetry. The model activates only 8B parameters when processing input (prefill) and 16B parameters when generating output (decode). Out of 552B total MoE parameters, each token lights up roughly 1.45% to 2.9% of the network.
Traditional autoregressive models use the same weight stack for both reading context and writing tokens. DeepSeek’s new Causal Encoder-Decoder architecture splits these into two pathways: a lightweight encoder for ingestion and a heavier decoder for generation.
Why This Works for Agent Workloads
Agent tasks have an extremely skewed token distribution. A typical agent call sends tens of thousands of tokens of context — system prompts, tool definitions, code repository prefixes, multi-turn history — and receives back dozens to hundreds of tokens of decisions or instructions. The input-to-output ratio can be 100:1 or even 1000:1.
By compressing the input pathway to 8B active parameters, DeepSeek slashed the cost of the most expensive part of agent inference. The 16B decode path ensures generation quality remains high where it matters most.
Benchmark Dominance: Flash Beats Pro
DeepSeek published benchmark comparisons showing V4.1-Flash ahead of V4 Pro and leading Western frontier models:
| Benchmark | V4.1-Flash | V4 Pro | Opus 5.0 | GPT-5.6 Sol |
|---|---|---|---|---|
| Terminal-Bench 2.1 | 90.6 | 87.9 | 89.1 | 88.8 |
| DeepSWE v1.1 | 74.2 | 62.7 | 74.0 | 73.0 |
| CyberGym | 88.1 | 83.3 | — | 84.5 |
| Terminal-Bench 3.0 | 30.0 | 11.8 | — | — |
| Terminal-Bench 4.0 | 31.2 | 12.4 | — | — |
The most striking gains are on Terminal-Bench 3.0 and 4.0, where V4.1-Flash roughly 2.5x to 3x the score of V4 Pro. On DeepSWE v1.1, the Flash model went from 54.4 (V4-Flash) to 74.2, leapfrogging both Opus 5.0 and GPT-5.6 Sol.
DeepSeek also reported wins on cybersecurity benchmarks, with V4.1-Flash claiming first place on CyberGym at 88.1%, ahead of GPT-5.6 Sol’s 84.5%.
KV Cache: 437x Compression Over Three Years
If asymmetric activation is the new trick, KV Cache compression is DeepSeek’s signature move. The progression tells a remarkable story:
| Model | Date | KV Cache per Token | Cumulative Compression |
|---|---|---|---|
| DeepSeek-V1 | Nov 2023 | 389,120 bytes | — |
| DeepSeek-V3.2 | Dec 2025 | 48,068 bytes | 8.1x |
| DeepSeek-V4-Flash | Apr 2026 | 3,514 bytes | 13.7x |
| DeepSeek-V4.1-Flash | Sep 2026 | 890 bytes | 437x total |
V4.1-Flash achieves this through three mechanisms working together:
- CSA2 (Compressed Sparse Attention 2): Cross-layer KV Cache reuse with Top-K indexing, operating in Full, Reindex, or Reuse modes. Redundant storage between layers is eliminated.
- FP4 KV Cache: The global cache is stored at 4-bit precision, down from the FP8 used in V4-Flash.
- DSPark speculative decoding: A lightweight module predicts tokens that the main model then confirms, boosting decode efficiency.
The hardware consequences are significant: compared to V4-Flash, V4.1-Flash needs only 1/4 the HBM and 1/8 the SSD storage. This is what makes the 1-million-token context window practical to serve at scale.
The Pricing Revolution: Up to 60% Cut
Architecture innovation translates directly into price cuts. DeepSeek adjusted its API pricing effective September 10:
| Pricing Tier | Previous (V4-Flash) | New (V4.1-Flash) | Change |
|---|---|---|---|
| Cache-hit input (off-peak) | ¥0.05/M tokens | ¥0.02/M tokens | -60% |
| Cache-miss input (off-peak) | ¥1.5/M tokens | ¥1.0/M tokens | -33% |
| Output (off-peak) | ¥4.5/M tokens | ¥4.0/M tokens | -11% |
In USD terms at off-peak rates: $0.003 per million cache-hit input tokens, $0.15 per million cache-miss input tokens, and $0.60 per million output tokens. Peak hours (Beijing time 9:00-12:00, 14:00-18:00) are 2x the off-peak rate.
The 60% cut on cache-hit input is not random. In agent workloads where system prompts and tool definitions repeat across calls, cache hit rates routinely exceed 99%. The cache-hit input line is the largest line item on the bill. DeepSeek targeted the deepest cut exactly where it hurts developers most.
V4 Pro Retirement: The Cheapest Tier Absorbs the Expensive One
DeepSeek confirmed that after September 14, 2026, all requests to deepseek-v4-pro will route to V4.1-Flash and be billed at V4.1-Flash rates. This continues until a future V4.1-Pro launch. The expensive tier is being folded into the cheap one.
For developers with hardcoded deepseek-v4-pro in production code, this is both a migration notice and a windfall — they get a better model at a lower price without changing anything.
Native Multimodal: Vision as a First-Class Citizen
V4.1-Flash ships with native visual understanding, not a bolted-on vision encoder. This is the culmination of DeepSeek’s multimodal push that began with V4-Flash-Vision-Exp in August 2026. The model can process screenshots, parse diagrams, inspect UI elements, and react to visual context within the same agent loop that handles code, search, and tool calls.
DeepSeek’s official partners — WorkBuddy (including CodeBuddy) and OpenCode — have already integrated V4.1-Flash, meaning developers can drop it into existing agent workflows immediately.
Open Source Under MIT License
Following its established pattern, DeepSeek published the V4.1-Flash weights on Hugging Face under the MIT license, accompanied by a technical report. This continues DeepSeek’s commitment to open-weight releases even as it prepares for a STAR Market IPO.
The company explicitly invited large-scale deployment partners with 2,000+ GPUs and storage clusters to collaborate on inference optimization — signaling that V4.1-Flash is designed for both API consumption and self-hosting at scale.
What This Means for DeepThink
The DeepThink reasoning engine sits at the core of the V4 model family. V4.1-Flash’s architecture improvements — asymmetric activation, CSA2, FP4 cache, DSpark — are infrastructure-level gains that benefit every reasoning task the engine performs. The 437x KV Cache compression means longer reasoning chains are cheaper to serve. The 1M-token context window means more complex multi-step problems fit in a single call.
For the broader industry, the message is clear: DeepSeek is not competing on parameter counts or raw compute. It is competing on architectural efficiency — squeezing more intelligence per dollar, per watt, per chip. And with V4.1-Flash, it has shown that the smallest model in a new architecture family can outperform flagships from both its own lineup and the West’s most prominent labs.
The V4.1 architecture is designed to scale to larger models. If Flash is this good, V4.1-Pro — whenever it arrives — will be the model to watch.