DeepSeek V4.1 Flash

DeepSeek V4.1 Flash is an efficiency-optimized Mixture-of-Experts (MoE) large language model released by DeepSeek on September 10, 2026, built on a new Causal Encoder-Decoder (CED) asymmetric architecture with 552B total parameters, activating only 8B for input and 16B for output, offering a 1 million token context window. The model natively supports multimodal visual understanding, offers high and xhigh reasoning levels, and compresses KV Cache to one-quarter of the previous generation. It surpasses V4-Pro and multiple flagship models on agent benchmarks including DeepSWE v1.1 (74.2%) and Terminal-Bench 2.1 (90.6%), with weights open-sourced under the MIT license.

The iCreat platform offers this model via the endpoint https://api.icreat.ai/llm/openai/v1/chat/completions (OpenAI Chat Completions compatible), at $0.3 per million input tokens, $1.2 per million output tokens, and $0.006 per million cached read tokens.

Model Positioning

DeepSeek V4.1 Flash is positioned as the smallest model in DeepSeek's new architecture family, designed for fast inference, high throughput, and cost efficiency.

The model uses an asymmetric CED architecture with 8B input activation and 16B output activation, delivering flagship-level intelligence at extremely low inference cost. In DeepSeek's official benchmarks, V4.1 Flash comprehensively surpasses V4-Pro across performance, cost, speed, and total runtime, making it the default model for the DeepSeek API. The model is suitable for coding assistants, conversational systems, and agentic workflows with high requirements for response speed and scalability.

On the iCreat platform, the model is served through an OpenAI Chat Completions compatible endpoint, and existing OpenAI SDKs work directly.

Core Capabilities

Asymmetric CED Architecture with Ultra-Low Inference Cost

552B total parameter MoE architecture with only 8B input and 16B output activation, using a new Causal Encoder-Decoder structure. KV Cache is just 890 bytes/token (HBM reduced to 1/4, SSD to 1/8), dramatically lowering agent cache-hit costs.

1 Million Token Ultra-Long Context

Supports 1,048,576 token context window with approximately 384K max output, enabling ingestion of entire codebases or complete document sets in a single request.

Native Multimodal Visual Understanding

Vision capability is integrated from the pre-training stage, natively supporting image input and understanding without a separate vision encoder.

Leading Agent and Coding Performance

DeepSWE v1.1 score of 74.2% (surpassing Claude Opus 5 at 74.0% and GPT-5.6 Sol at 73.0%), Terminal-Bench 2.1 at 90.6%, Automation-Bench at 54.8% — leading in agent and coding benchmarks.

Streaming Output and Function Calling

Supports streaming output (SSE), function calling, structured output, context caching, and multi-turn conversations across the Chat Completions endpoint.

Pricing

Note: The above are iCreat platform prices. DeepSeek's official pricing uses peak/off-peak time-based billing (off-peak: input $0.15, output $0.60, cache read $0.003; peak: 2x off-peak). iCreat uses flat pricing regardless of time. The iCreat platform price shall prevail.

Application Scenarios

• Coding assistants and repository-scale development tasks
• Agentic workflows and multi-step tool calling
• Long-document analysis and large-scale codebase understanding
• Multimodal document processing (image, chart understanding)
• Enterprise conversational systems and customer service knowledge bases

Model Comparison

Comparison Table 1: DeepSeek V4.1 Flash vs V4-Pro

Note: V4.1 Flash comprehensively surpasses V4-Pro across performance, cost, speed, and total runtime, and has been set as the default DeepSeek API model; V4-Pro is being phased out.

Comparison Table 2: DeepSeek V4.1 Flash vs Same-Tier Agent Models

Note: DeepSeek V4.1 Flash is the lowest-priced of the three, ranks first on DeepSWE, and is the only MIT-licensed model.

Why Choose DeepSeek V4.1 Flash?

• Asymmetric CED architecture with only 8B input activation for ultra-low inference cost
• 1 million token ultra-long context with KV Cache compressed to 1/4 of the previous generation
• DeepSWE v1.1 score of 74.2%, surpassing Claude Opus 5 and GPT-5.6 Sol
• Lowest price among comparable models ($0.3/$1.2), far below Gemini 3.7 Flash and Sonnet 5
• MIT open-source license for self-hosting
• Native multimodal visual understanding without a separate vision encoder

Specifications

Architecture

The iCreat platform's DeepSeek V4.1 Flash is served through an OpenAI Chat Completions compatible endpoint, allowing users to call it with any OpenAI API-compatible SDK. Requests must include the API Key in the Authorization header. The request body contains the model field (deepseek-v4.1-flash), a messages array (system/user/assistant messages), and optional thinking object (to enable thinking mode), reasoning_effort field (reasoning intensity), stream parameter (streaming output), and standard parameters such as temperature/max_tokens. The model uses a new Causal Encoder-Decoder asymmetric architecture with 40 transformer layers split into a 20-layer causal encoder (processing input context) and a 20-layer decoder (generating output tokens). KV Cache is only 890 bytes/token, dramatically reducing cache costs for long-context and multi-turn agent scenarios. The model supports context caching, with cached read pricing at one-fiftieth of the standard input rate.

Notes

• The model field must use the iCreat platform's model_code (deepseek-v4.1-flash), not the vendor's original model name
• The model supports high and xhigh reasoning levels, with xhigh as the maximum reasoning intensity
• Cache read costs only $0.006/M tokens — one-fiftieth of the standard input rate, significantly reducing costs for agent scenarios
• The model natively supports image input without a separate vision encoder
• Weights are open-sourced under the MIT license; self-hosting requires approximately 300GB GPU memory (4-bit quantization)
• Please safeguard your API Key and avoid hardcoding it in client-side code or public repositories

Frequently Asked Questions

How is the billing calculated?​

Billing is based on actual token usage. Input is $0.3/M tokens, output is $1.2/M tokens, and cached read is $0.006/M tokens. For example, 1M input + 0.5M output costs $0.3 + $0.6 = $0.9. The iCreat platform price shall prevail. DeepSeek's official pricing uses peak/off-peak time-based billing; iCreat uses flat pricing regardless of time.

What is the difference between DeepSeek V4.1 Flash and V4-Pro?​

V4.1 Flash uses a new CED asymmetric architecture (552B total parameters, 8B input / 16B output activation), comprehensively surpassing V4-Pro across performance, cost, speed, and total runtime. V4.1 Flash has been set as the default DeepSeek API model; V4-Pro is being phased out. Both are open-sourced under the MIT license.

How large is the context window?​

Supports 1 million tokens (1,048,576 tokens) context window with approximately 384K max output. KV Cache is only 890 bytes/token — one-quarter of the previous generation — dramatically reducing cache costs for long-context scenarios.

Is thinking mode supported?​

Yes. Add "thinking": {"type": "enabled"} to the request body. The model supports high and xhigh reasoning levels, with xhigh as the maximum reasoning intensity, suitable for complex reasoning tasks.

Does it natively support image input?​

Yes. Vision capability is integrated from the pre-training stage, natively supporting image input and understanding without a separate vision encoder. Images can be passed via the image_url type in the messages content array.

Is the model open source?​

Yes. Model weights are open-sourced under the MIT license on Hugging Face, permitting commercial use, modification, and redistribution. Self-hosting requires approximately 300GB GPU memory (4-bit quantization); a multi-GPU server is recommended.