Introduction
On September 10, 2026, DeepSeek launched V4.1 Flash, a new multimodal Mixture-of-Experts (MoE) model released under the MIT license. The model totals 552B parameters and adopts a brand-new Causal Encoder-Decoder (CED) asymmetric architecture. During the inference phase, it activates only 8B parameters for input prefill and 16B parameters for token generation. It supports a 100 million-token context window and native multimodal image capabilities. According to official technical reports, its performance on many Agent benchmarks exceeds DeepSeek V4 Pro, and some test results beat Opus 5 and GPT-5.6 Sol.
DeepSeek optimized the global KV Cache compression to 890 bytes per token. This change cuts HBM memory demand to one-quarter of the previous V4 Flash generation, while SSD storage requirements drop to 1/8. The newly introduced API pricing tier brings off-peak pricing at half of peak rates, with a concurrency cap raised to 2500. Starting September 14, 2026, all API requests originally routed to V4 Pro will automatically redirect to V4.1 Flash and be billed at V4.1 Flash rates. This article disassembles the structural innovations, benchmark results, pricing adjustments and access approaches of DeepSeek V4.1 Flash, and analyzes its impact on domestic large model marketplaces.
1. What is DeepSeek V4.1 Flash
DeepSeek summarizes V4.1 Flash in three keywords: stronger, faster, and more affordable. It is the smallest-sized model within the new structural series. Its core engineering theme is *Pushing the Limits of KV Cache Compression and Reasoning*. The official technical report directly targets extreme compression for KV Cache, which has long become a primary memory bottleneck for long-context Agent workloads.
Key facts released alongside the launch:
- The Hugging Face repository
deepseek-ai/DeepSeek-V4.1-Flashwas created at 02:17 UTC on September 10, 2026, containing a 51-page technical report, prompt encoding references and PyTorch implementation code. - Configuration files show 40 layers and 384 routing experts. Each token activates 6 experts plus 1 shared expert, with a maximum vocabulary size of 1,048,576 and FP4 expert weight storage.
- Tencent WorkBuddy, CodeBuddy and OpenCode have fully integrated the model as official cooperative partners.
- The Hacker News post titled *DeepSeek launching v4.1 flash cheaper and more capable than v4 pro* earned 399 upvotes on September 9.
2. Asymmetric Architecture: Causal Encoder-Decoder with 8B Input /16B Output
Traditional Decoder-only models reuse the same parameter set for both prefill (prompt reading) and decode (token-by-token generation). For long-context Agent scenarios, the prefill stage loads extensive tool outputs, documents and historical dialogue records. The computation and KV Cache storage overhead of prefill gradually becomes the major cost bottleneck.
DeepSeek’s CED architecture separates the two stages and uses different activated parameter scales:
| Stage | Activated Parameters per Token | Responsibilities |
|---|---|---|
| Prefill (Input Encoding) | 8B | Load prompts, tool return content, documents and images |
| Decode (Output Decoding) | 16B | Generate reasoning logic, code and text replies |
The official document states that asymmetric input and output design substantially reduces cost compared to models of similar parameter scale. Combined with large-scale reinforcement learning post-training, V4.1 Flash outperforms many flagship models including DeepSeek V4 Pro on multiple benchmarks.
Beyond CED, the technical report lists a series of supporting innovations: Compressed Sparse Attention 2 (CSA2), cross-layer KV and index reuse, layered sparse retrievers, Single-Pass mHC, Engram n-gram lookup tables, DSpark forward routing and FP4 primary KV Cache. The pre-training corpus covers 45T tokens of multimodal data.
3. KV Cache Compressed to 890 Bytes per Token: HBM Requirement Cut to 1/4, SSD to 1/8
The global KV Cache of V4.1 Flash takes only 890 bytes for each token, roughly one quarter of V4 Flash and 1/437 of early DeepSeek models. DeepSeek splits KV Cache into two categories and optimizes them separately:
- Global KV Cache (Resident in HBM): Cross-layer reuse via CSA2 and FP4 quantization, compressed down to 890 bytes per token, 1/4 of V4 Flash.
- Persistent KV Cache (Resident in SSD or host memory): Optimized through SWA Bounded Replay deployment, reduced to around 1/8 of V4 Flash.
The official explanation highlights the direct financial impact: cached input accounts for a large proportion of Agent workload billing. The heavy compression of KV Cache greatly reduces the running cost of Agent tasks. Under the new pricing rule, cached hit input costs 0.04 yuan per million tokens during peak hours, which is 1/50 of uncached input price. This compression delivers a direct reduction in inference expense.
4. Benchmark Performance: Agent Tasks Surpass V4 Pro Across the Board
The table below compares V4.1 Flash against Opus 5, GPT-5.6 Sol, Kimi K3, GLM-5.3, V4 Pro and original V4 Flash under maximum reasoning settings, extracted from Table 3 of the technical report.
| Benchmark | Opus 5 | GPT-5.6 Sol | Kimi K3 | GLM-5.3 | V4 Pro | V4 Flash | V4.1 Flash |
|---|---|---|---|---|---|---|---|
| GPQA Diamond | 93.4 | 94.1 | 92.9 | 88.1 | 92.4 | 89.9 | 90.9 |
| Codeforces Rating | — | — | — | — | 3348 | 3289 | 3471 |
| MathArena Apex | — | — | 65.6 | — | 65.8 | 58.6 | 65.6 |
| Terminal-Bench 2.1 | 89.1 | 88.8 | 88.3 | 88.2 | 87.9 | 82.7 | 90.6 |
| DeepSWE v1.1 | 74.0 | 73.0 | 67.5 | 66.9 | 62.7 | 54.4 | 74.2 |
| Automation-Bench | 50.3 | 45.8 | 46.7 | 48.8 | 43.2 | 37.7 | 54.8 |
| Agents' Last Exam | 28.6 | 26.7 | 27.6 | 28.5 | 25.7 | 25.2 | 31.8 |
| CyberGym | — | 84.5 | 80.0 | 84.5 | 83.3 | 76.7 | 88.1 |
| Terminal-Bench 4.0 | 51.8 | 39.9 | 12.6 | 37.9 | 12.4 | 7.0 | 31.2 |
| HLE | 56.3 | 44.5 | 43.5 | 42.0 | 42.7 | 37.8 | 36.8 |
Three critical takeaways from benchmark data:
- Agent programming becomes its strongest suit. DeepSWE v1.1 jumps from 54.4 in V4 Flash to 74.2, exceeding Opus5’s 74.0. Terminal-Bench 2.1’s score of 90.6 ranks highest among these seven models.
- Pure reasoning gaps remain. HLE score at 36.8 sits clearly below Opus 5’s 56.3, and Terminal-Bench 4.0 at 31.2 still has a noticeable gap against Opus 5’s 51.8.
- Stable cross-framework performance. The report tested DeepSWE v1.1 across six harness platforms including Claude Code, Codex, OpenCode, Pi, mini-SWE and DeepSeek Harness, achieving scores ranging from 65.5 to 74.2. This proves the model capability is not bound to a single evaluation harness.
The report also publishes the impact of reasoning effort adjustment. Raising effort value from 25 to 100 lifts average Pass@1 of eight benchmark tasks from 67.1% to 76.3%, while output token volume increases roughly 2.5 times. When the parameter sits between 60 and 80, most precision is retained with less than half the token budget. The public API provides three tiers: low / high / max, corresponding to effort 50 /75 /100.
5. API Pricing and Model Routing Changes
The new pricing scheme took effect at 12:00 on September 10, 2026, with off-peak rates set at half of peak prices.
DeepSeek V4.1 Flash API Price Table (Unit: Yuan / Million Tokens)
| Time Window | Input (Cache Hit) | Input (Cache Miss) | Output |
|---|---|---|---|
| Off-Peak Hours | 0.02 | 1.0 | 4.0 |
| Peak Hours | 0.04 | 2.0 | 8.0 |
Peak window: Working days 9:00–12:00 and 14:00–18:00 Beijing Time. All remaining time including weekends and holidays counts as off-peak.
For comparison, DeepSeek V4 Pro original pricing was 0.30 yuan (peak) /0.15 yuan (off-peak) for cache hit input, 9 yuan /4.5 yuan for cache miss input, and 27 yuan /13.5 yuan for output. V4.1 Flash supports a maximum context of 1M tokens and maximum generation length of 384K, with concurrency limit lifted to 2500, while V4 Pro had no concurrency control.
Key routing rules:
- Calling the latest unified model ID
deepseek-flashroutes traffic to V4.1 Flash. - Legacy aliases
deepseek-v4-flash,deepseek-v4-flash-vision-exptemporarily point to V4.1 Flash. The old V4 Flash and V4 Flash Vision Exp models have been retired. - Starting at 12:00 on September14, all requests to
deepseek-v4-prowill be redirected to V4.1 Flash and charged under Flash pricing until V4.1 Pro becomes available.
DeepSeek’s official justification for the retirement: V4.1 Flash outperforms V4 Pro comprehensively in performance, cost, speed and total latency. From an output cost perspective, migrating from V4 Pro to V4.1 Flash cuts billing expenses to approximately 30% of the original.
6. How to Access DeepSeek V4.1 Flash
DeepSeek API supports both OpenAI and Anthropic compatible formats. Developers only need to specify the model name deepseek-flash in request parameters.
For OpenAI compatible endpoints, developers can reuse existing OpenAI SDKs and switch base URL and API key to connect. The Anthropic compatible base URL is https://api.deepseek.com/anthropic, which can run Claude Code style tool calling workflows.
For self-host deployment, the official HuggingFace repository provides minimal PyTorch inference implementation and weight conversion scripts. The official statement shows full support for open-source community inference adaptation. Running full-scale deployment requires GPU clusters with 24G VRAM and storage arrays; consumer-grade hardware cannot load the complete model weights.
When developers manage multiple LLM model endpoints in production, an API gateway can simplify unified routing, authentication and traffic monitoring. 4sapi delivers such gateway capabilities to streamline multi-model API management.
7. Impacts on Domestic Large Model Marketplaces
Within a few days, the new pricing and routing policy will propagate to third-party model marketplaces. Platform operators and enterprise developers should focus on three major points for model selection and integration:
- Model ID Alignment. DeepSeek has adopted the ID
deepseek-flashwithout version suffixes. Third-party platforms usually retain model IDs with timestamps. When migrating, teams need to compare cache hit rates, peak/off-peak billing rules, and concurrency caps. - V4 Pro routing transition. After September14, requests targeting
deepseek-v4-prowill be mapped to V4.1 Flash. If your business logic relies on unique V4 Pro behaviors, regression tests must finish before the deadline. - Peak/off-peak scheduling. Off-peak half-price billing brings obvious economic benefits for batch processing, offline evaluation and night-running Agent workloads. Task scheduling pipelines can add time window judgement logic to cut inference cost.
8. Frequently Asked Questions
What are the differences between V4.1 Flash and V4 Flash?
The core upgrade replaces Decoder-only architecture with Causal Encoder-Decoder. Prefill activated parameters drop to 8B. KV Cache is compressed to roughly 1/4 of the prior generation. Native image capability is integrated, removing the requirement for separate Vision Exp model calls. Agent benchmark scores have substantial gains: DeepSWE v1.1 from 54.4 to74.2, Terminal-Bench 2.1 from82.7 to90.6.
Why is V4 Pro retired?
Official test data shows V4.1 Flash surpasses V4 Pro in performance, cost, speed and end-to-end latency. V4.1 Flash scores higher in most Agent benchmarks, though pure reasoning benchmarks like GPQA Diamond and HLE remain slightly lower. After September14, all V4 Pro API traffic is routed to V4.1 Flash, pending V4.1 Pro release.
What is the difference between the 552B parameter count and the 485B number displayed on Hugging Face?
The technical report and official release use 552B as the total back-bone parameter count. The 485B figure on Hugging Face comes from Safetensors metadata auto-statistics, counting FP8 / FP4 mixed precision weight and component statistics, while the paper takes the theoretical parameter value as the standard.
How does off-peak pricing work?
Peak hours are 9:00–12:00 and14:00–18:00 Beijing time on workdays. All other periods including weekends are treated as off-peak, and all token unit prices are halved. Billing is calculated based on the timestamp when requests complete.
Can it run locally?
The model can be downloaded locally under the MIT license with minimal PyTorch inference code. However, the 552B total parameters paired with a 100 million-token context window impose heavy storage and memory pressure. The official notes that large-scale deployment requires GPU clusters with 24G VRAM and dedicated storage pools, and individual end-user devices are not practical for local inference.
Conclusion
DeepSeek V4.1 Flash uses asymmetric architecture and extreme KV Cache compression to deliver stronger capability at drastically reduced pricing, while phasing out its former flagship V4 Pro. This type of major model upgrade paired with price reduction is relatively rare in recent large model evolution cycles. All data in this article comes from DeepSeek official announcements, V4.1 Flash technical report, API pricing page and HuggingFace repository materials published on September 10, 2026. All pricing and routing rules are subject to real-time updates from the official website.
The combination of half-price off-peak billing and high-performance Agent capability reshapes the cost baseline for long-context autonomous agent services. For enterprise teams running continuous batch evaluation and automated coding agents, reasonable scheduling around peak/off-peak windows can bring substantial inference savings. When managing multi-vendor model access in production environments, unified gateway services reduce the operational overhead of maintaining multiple API endpoints.
International access: https://4sapi.com
Domestic access: https://4sapi.cn
