DeepSeek-V4.1-Flash
by deepseek-ai
DeepSeek's open multimodal MoE: images and text in, 1M-token context, 8B parameters active per token in prefill
deepseek-ai/DeepSeek-V4.1-FlashDeploy DeepSeek-V4.1-Flash
Single-tenantMixpeek has no managed extractor for this model. On a single-tenant deployment you upload the weights and a custom plugin serves them next to the rest of your pipeline.
Overview
DeepSeek-V4.1-Flash is an open-weight multimodal Mixture-of-Experts model from DeepSeek, released in September 2026 under the MIT license. It reads images and text natively, writes text, and takes contexts of up to one million tokens. Its 552B backbone activates 8B parameters per token while reading input and 16B while generating, which DeepSeek describes as built for input-heavy agent workloads.\n\nThe card reports a global KV cache per token about four times smaller than DeepSeek-V4-Flash. On the base model it lists 95.6 on DocVQA, 86.0 on RefCOCO, 77.9 on CVBench and 56.5 on MMMU-Pro.
Architecture
Causal encoder-decoder Transformer: 40 layers split into a 20-layer causal encoder and a 20-layer decoder, with the decoder's global KV cache projected from the encoder's final hidden states. MoE layers have 1 shared and 384 routed experts, 6 routed experts active per token. Compressed Sparse Attention 2, Engram conditional memory (196B parameters, sparse lookup) and DSpark speculative decoding. Images go through DeepSeek-ViT (trained from scratch, 2D-RoPE, 3x3 pixel-unshuffle) and a two-layer MLP projector, trained jointly with text from the start of pre-training on 45T tokens.
Mixpeek SDK Integration
# DeepSeek-V4.1-Flash runs on your own multi-GPU servers with the inference/
# code in its repository; it ships no chat template, so prompts go through the
# encoding/ reference implementation. Store what it writes about each page or
# image as a text object, so a Mixpeek text collection can search it.
import requests
for page in described_pages: # e.g. {"key": "contracts/acme.pdf#page=4", "text": "Termination clause ..."}
requests.post(
"https://api.mixpeek.com/v1/buckets/bkt_your_bucket/objects",
headers={"Authorization": "Bearer API_KEY", "X-Namespace": "ns_your_namespace"},
json={"blobs": [{"property": "description", "type": "text", "data": page["text"]}],
"metadata": {"source_key": page["key"]}},
)Capabilities
- Image and text input, text output
- Contexts up to 1M tokens
- 8B active parameters per token in prefill, 16B in decode
- Controllable reasoning effort from 1 to 100
- MIT license
Use Cases on Mixpeek
Benchmarks
| Dataset | Metric | Score | Source |
|---|---|---|---|
| DocVQA | LLM-Judge, 4-shot | 95.6 | Model card: deepseek-ai/DeepSeek-V4.1-Flash (self-reported, base model) |
| RefCOCO-avg | [email protected], 0-shot | 86.0 | Model card (self-reported, base model) |
| CVBench | EM, 4-shot | 77.9 | Model card (self-reported, base model) |
| MMMU-Pro | EM, 4-shot | 56.5 | Model card (self-reported, base model) |
| MMLU-Pro | EM, 5-shot | 74.1 | Model card (self-reported, base model; V4-Flash-Base 68.3) |
Performance
A 552B-parameter model needs a multi-GPU deployment even with 8B to 16B active per token. The card recommends temperature 1.0, top_p 0.95 and max_tokens of 256K or more. We have not measured it.
Common Pipeline Companions
Frequently Asked Questions
What is DeepSeek-V4.1-Flash?
An open-weight multimodal Mixture-of-Experts model from DeepSeek that reads images and text, writes text, and takes up to one million tokens of context. It is MIT licensed.
How many parameters does DeepSeek-V4.1-Flash use per token?
Its backbone has 552B parameters, and it activates 8B per token while reading input and 16B while generating, according to the model card.
How well does it read documents and images?
The card reports 95.6 on DocVQA, 86.0 on RefCOCO, 77.9 on CVBench and 56.5 on MMMU-Pro for the base model. These are DeepSeek's own numbers.
Can I search what DeepSeek-V4.1-Flash reads with Mixpeek?
Yes. Run the model on your side, store what it writes about each page or image as a text object with the source in metadata, and search those with a text retriever, as in the example on this page.
Specification
Research Paper
DeepSeek-V4.1-Flash: Pushing the Limits of KV Cache Compression
arxiv.orgBuild a pipeline with DeepSeek-V4.1-Flash
Add this model to a processing pipeline alongside other extractors. Combine with retrieval stages for end-to-end search.
Run it on your own data, free