DeepSeek-V4.1-Flash: an open-weights giant for pennies. Architecture, performance, pricing and self-hosting

01DeepSeek in 2026: where V4.1-Flash comes from

DeepSeek is a private company from Hangzhou, China, and a subsidiary of the investment fund High-Flyer Capital Management. It released its first model, DeepSeek-Coder, on November 2, 2023, and V4.1-Flash is its 23rd release. (Mungomash)

The last few months of the V4 line

Why DeepSeek is focused on the KV cache

Agents run longer and with huge inputs, so tasks are increasingly “input heavy.” According to DeepSeek's technical report, the main obstacle to cheaper inference today is compute-heavy input processing (prefill) and large KV caches, which strain HBM memory, SSDs and bandwidth. V4.1-Flash is designed to attack exactly that bottleneck. (Hugging Face Papers)

DeepSeek describes the model as a new starting point for further scaling of the architecture, pre-training and post-training. (arXiv)

02Architecture deep dive

V4.1-Flash combines MoE with a new causal encoder-decoder and five tricks to shrink the KV cache. The result is 890 bytes of KV cache per token, roughly a quarter of V4-Flash and 437x less than the original DeepSeek V1. (Hugging Face: DeepSeek)

1. Mixture-of-Experts: 552 billion, but only a fraction at work

Each MoE layer has 1 shared and 384 routed experts; 6 routed experts are activated per token. The router picks only a small, relevant part of the network for each token. (Hugging Face: DeepSeek) For performance planning, the active parameter count therefore matters more than the total. (MindStudio)

2. Causal encoder-decoder (CED)

The network has 40 layers: a 20-layer causal encoder followed by a 20-layer decoder. The decoder's global KV cache is not computed from each of its layers separately; it is projected from the encoder's final states. As a result, the model activates only 8 billion parameters when processing input and 16 billion when generating, which mainly makes agentic tasks with long inputs cheaper. (Hugging Face: DeepSeek)

3. Compressed Sparse Attention 2 (CSA2)

Each attention layer has one fixed mode out of three: Full, Reindex or Reuse. Layers share the main KV and indexer keys and reuse sparse attention indices. In the decoder, a hierarchical sparse indexer also limits deeper layers to candidates picked by the first Full layer, so indexing cost does not grow with context length. The first two encoder layers use sliding window attention (SWA). (Hugging Face: DeepSeek, arXiv)

4. FP4 KV cache

The main KV cache is stored in the 4-bit E2M1 format with one E4M3 scaling factor per 16 channels. Together with CSA2, that gives the 890 bytes per token mentioned above. (Hugging Face: DeepSeek)

5. SWA Bounded Replay

Missing sliding window states are not reconstructed from disk but by replaying only the most recent tokens. Persistent KV cache (on SSD or in host memory) drops to roughly an eighth of V4-Flash's. (Hugging Face Papers)

Other components

• Single-Pass mHC: modified residual stream mixing with a fast Mega-mHC kernel.
• Engram: a conditional memory with 196 billion parameters that the model accesses sparsely per token.
• DSpark: speculative decoding with semi-autoregressive drafting and confidence-based verification.

Source: Hugging Face: DeepSeek

Training

The model was pre-trained on a 45 trillion token multimodal corpus and went through extensive post-training for text and multimodal agentic scenarios. (Hugging Face Papers)

03Specifications

04Benchmarks and independent testing

V4.1-Flash excels at terminal tasks, long coding work and finding vulnerabilities, but falls well behind the frontier on the general agentic Terminal-Bench 4.0.

Results according to DeepSeek

Where it leads and where it loses

On CyberGym (88.1%) it beat every model that has a result in DeepSeek's table; Claude Opus 5 has no number there. (Hugging Face: DeepSeek) For comparison, the specialized Gemini 3.8 Flash Cyber scores 86.2% on the same test.

Hands-on tests

Independent testers ran into weaknesses outside the benchmarks. In a Rubik's cube simulation, the colors changed incorrectly during moves, and an attempt to copy an image in the style of Microsoft Paint produced a stylized but low-detail result. There is a gap between strong numbers and quality on creative and spatial tasks. (MindStudio)

05How it compares with the September 2026 models

V4.1-Flash is among the best at long coding work, and on price only Xiaomi's MiMo-V2.6-Pro, the only other open-weights model in the table, competes with it. On general agentic tasks it sits in the lower half.

On Terminal-Bench 2.1, V4.1-Flash scores 90.6%, more than Gemini 3.8 Flash (89.4%), Xiaomi MiMo-V2.6-Pro (89.9%) and Muse Spark 1.3 (88.8%).

For more on the competitors, see our GPT-6 Astra guide, Claude Fable 5.1 guide, Gemini 3.8 Flash guide, Grok 4.7 guide, Muse Spark 1.3 deep dive and article on Xiaomi MiMo V2.6 Pro.

For all seven September models side by side, with pricing and benchmark charts, see our big comparison of September 2026 AI models.

06API pricing: peak, off-peak and a price cut

At peak, V4.1-Flash costs $0.30 per million input tokens and $1.20 per million output tokens; off-peak, exactly half. It was the only September model to cut both input and output prices; Claude Fable 5.1 cut only its cache read price. (Capital and Compute)

Compared with V4-Flash, that is a 32% cut on input, 9% on output and 57% on cache reads. (Capital and Compute) For perspective: GPT-6 Astra and Claude Fable 5.1 cost $10 per million input tokens, 33x more than V4.1-Flash at peak.

The V4-Pro retirement saga

At launch, DeepSeek announced that from 4:00 UTC on September 14, every request to deepseek-v4-pro would be answered by V4.1-Flash. It later reversed the decision, and V4-Pro keeps running at its existing prices. (DeepSeek: changelog) The older V4-Flash and V4-Flash-Vision-Exp, however, were retired from the direct API and their names now redirect to V4.1-Flash, so check model IDs and prices with resellers separately.

Cost tip

Run batch and non-urgent jobs (overnight analysis, data generation) off-peak at half price, and keep a stable prompt prefix to benefit from the $0.006 cache price.

07Open weights in practice: self-hosting

You can download V4.1-Flash from Hugging Face (deepseek-ai/DeepSeek-V4.1-Flash) and run it commercially under the MIT license. “Open” does not mean “free,” though: you need either your own compute or a paid API. (MindStudio)

What drives the hardware requirements

• Weight storage must cover all 552 billion parameters, even though only a fraction is used per token. The experts have to live somewhere fast enough: in VRAM, in RAM with offloading or on NVMe.
• Active parameters (8B and 16B) determine compute and throughput per token; for speed planning they matter more than the total.
• Quantization is a must for smaller hardware. DeepSeek has not published official minimum VRAM requirements.

Source: MindStudio

Community variants

Realistic paths

1. Rented cloud GPUs with DeepSeek's official inference code: the fastest start.
2. A workstation with two professional GPUs and a heavily quantized version: for teams that want to keep data in-house.
3. A third-party hosted API: no hardware of your own and no servers in China. NVIDIA's trial endpoint on build.nvidia.com is for testing only: its terms prohibit confidential and personal data, and usage is logged.

At launch, most local runs were on multi-GPU servers or in the cloud, not on gaming rigs. (MindStudio)

08Quick start with the API

DeepSeek's direct API uses the ID deepseek-flash and a 1 million token context. (Mungomash) The DeepSeek API is compatible with the OpenAI format, so you only need to change the base URL and model:

import os
from openai import OpenAI

client = OpenAI(
    base_url="https://api.deepseek.com",  # check the DeepSeek documentation
    api_key=os.environ["DEEPSEEK_API_KEY"],
)

response = client.chat.completions.create(
    model="deepseek-flash",
    messages=[
        {"role": "user", "content": "Go through this log and find the cause of the error."}
    ],
)

print(response.choices[0].message.content)

Check the base URL and reasoning parameters in the DeepSeek API documentation before you deploy. For self-hosting, use your own server's endpoint (vLLM and similar tools offer an OpenAI-compatible interface) and the model deepseek-ai/DeepSeek-V4.1-Flash.

09Data, jurisdiction and security

With DeepSeek, the data decision is simpler than with closed models: if you do not want to send data to China, you can host the model yourself.

Three ways to run it, from a data perspective

DeepSeek is based in Hangzhou, China. (Mungomash) Check the specific data processing terms of the direct API in DeepSeek's policies, and consult a lawyer for sensitive data; the table above only describes where the data physically ends up.

Model security

• Strong at finding vulnerabilities: 88.1% on CyberGym, more than GPT-5.6 Sol (84.5%) and the specialized Gemini 3.8 Flash Cyber (86.2%).
• Weaker at building exploits: 15.3% on ExploitGym versus 33.7% for GPT-5.6 Sol and 22.1% for Claude Opus 5.
• Open weights mean no provider safeguards. Unlike Claude Fable or GPT-6 Astra, self-hosting has no central filter that would reroute risky requests. The operator is responsible for the usage rules.

Result sources: Hugging Face: DeepSeek, NVIDIA, our Gemini 3.8 Flash guide (Cyber)

10Decision guide

V4.1-Flash is the default choice when cost or control over data decides; for the hardest general agents, look elsewhere.

11DeepSeek-V4.1-Flash FAQ

12Verdict

DeepSeek-V4.1-Flash delivers top coding results for pennies, plus open weights under the MIT license. For cheap agentic coding, long inputs and companies that want to keep data in-house, it is one of the strongest candidates. Its closest open rival is Xiaomi MiMo V2.6 Pro, which has a higher overall Artificial Analysis score (46 vs 39 points) but trails on DeepSWE (71.9% vs 74.2%).

It is not enough for general agents, computer use and creative tasks, though, and the direct API leads to China. Want to deploy AI in your company and choose the right model? We can help as part of our AI implementation service.

13Sources

• Hugging Face: DeepSeek-V4.1-Flash (official model card)
• arXiv: DeepSeek-V4.1-Flash technical report
• Hugging Face Papers: report abstract
• NVIDIA: DeepSeek-V4.1-Flash model card
• DeepSeek API: updates
• DeepSeek API: pricing
• DeepSeek: V4.1-Flash announcement
• MindStudio: how to run V4.1-Flash locally
• Capital and Compute: September 2026 models and prices
• Mungomash: frontier model overview
• Local AI Zone: September 2026 model updates
• TeamDay: best AI models of 2026
• Hugging Face: NVFP4 from NVIDIA
• Hugging Face: NVFP4 conversion (s-zaizen)
• Hugging Face: GGUF (vcruz305)
• Hugging Face: EXL3 for 2x RTX PRO 6000