DeepSeek V4.1 Flash Open-Source: 552B Architecture, Pricing, and Migration

What is DeepSeek V4.1 Flash?

DeepSeek V4.1 Flash was released on September 10, 2026, with open weights under MIT license. The official API model name is deepseek-flash. It is a native multimodal MoE that accepts text and images and generates text, using an asymmetric Causal Encoder-Decoder (CED) structure: 552B total backbone parameters, with roughly 8B activated per input token during prefill and 16B activated per output token during decode.

The CED backbone has 40 layers (20 encoder + 20 decoder). The decoder's global KV cache is projected from the encoder's final hidden state rather than generated per layer, which lets input and output stages use different compute budgets.

How the 552B / 8B / 16B numbers relate

MoE models don't activate every parameter for every token. 552B is the total backbone size; 8B and 16B are the active parameters at the input and output stages respectively, not three interchangeable model sizes. Don't assume deployment only needs to load 8B of weights.

Agent benchmark gains vs V4 Pro

V4.1 Flash improves on code, terminal, and automation benchmarks but trails V4 Pro on GPQA Diamond. Agent capability is clearly better, not a uniform upgrade. Benchmark results depend on model version, agent framework, and inference budget, so reproduce with your own tasks before drawing conclusions.

Official pricing and how to estimate cost

Pricing tiers reflect off-peak and peak windows (peak: Mon-Fri 09:00-12:00 and 14:00-18:00 Beijing time). Cached input, uncached input, and output are billed separately; a 50x ratio between cached and uncached input doesn't make the whole task 50x cheaper. Check our /model-pricing page for current gateway rates, which may differ from upstream direct pricing.

Example bill calculation

For a workload of 8M cached input tokens, 2M uncached input tokens, and 0.5M output tokens in a single window, total cost equals (cached millions x cached rate) + (uncached millions x uncached rate) + (output millions x output rate). Plug in the off-peak and peak rates from the official pricing page to compare.

Why KV cache compression matters for long sessions

For agents that repeatedly read the same codebase, system docs, or conversation history, prefix caching avoids recomputation. V4.1 Flash compresses global KV cache to 890 bytes/token, roughly a quarter of V4 Flash. Note that official numbers describe cache resources, not total VRAM, weights, or cluster capacity. Keep reusable prefixes stable across turns and verify actual cache hits via usage fields rather than assuming similarity equals a hit.

Calling the API

New integrations should use the model name deepseek-flash. The official Chat Completions endpoint accepts a standard request body with model, messages, and optional parameters like max_tokens, stream, thinking, and reasoning_effort. Responses arrive in JSON with the answer in choices[0].message.content; always verify HTTP status, content, and the usage block. The model understands images but does not generate them; follow the official image input guide to embed image content blocks. When routing through a third-party gateway, confirm the base URL, model ID, authentication, and protocol support.

Legacy model names and the V4 Pro switch

Keeping an old alias does not lock in the old weights. Migrate in three steps: export current model names and sample requests, revalidate with deepseek-flash across text, image, tools, and caching, then update client defaults and reconcile billing. For reproducible tasks, record call date, actual provider, and response payload, not just the alias.

DeepSeek Harness v0.1.5

Harness is the agent runtime that connects the model to files, commands, and tools. v0.1.5 adds image and PDF upload, workspace file tree with artifact preview, improved long-session resume, bidirectional parent/child agent communication, queued messages, task intervention, and experimental Agent Teams (off by default). Launch with npx @deepseek-ai/dsh web on a system with Node.js installed; note that npx fetches the latest available package, not necessarily v0.1.5, so pin the version when reproducibility matters.

Can it run on a regular workstation?

No. The 8B activation figure does not imply 8B-class hardware suffices; the backbone is still 552B, and deployment depends on weight precision, runtime, cache, concurrency, and storage. The MIT-licensed weights and tech report are available from the official repo. Large-scale deployment partnerships target teams with roughly 2000 GPUs and cluster storage, a collaboration requirement, not a published minimum config. Teams focused on applications or agent validation should gather quality and cost data via the API before evaluating self-hosting.

FAQ

550B or 552B?

Official release pages and the model card state 552B for the backbone. 550B is a rounded paraphrase; use 552B for specs and deployment estimates.

What is the API model ID?

Use deepseek-flash. Do not construct deepseek-v4.1-flash yourself; a third-party alias does not make it an official ID.

Does the cheapest cached-input rate apply to all tokens?

No. That rate only covers off-peak cache hits on input. Uncached input, output, and peak-window traffic have separate pricing.

Does V4.1 Flash generate images?

No. It accepts text and images as input and generates text only. Use it for screenshot understanding, chart analysis, and similar tasks.

Is V4 Pro already retired?

Not as of September 10, 2026. The official switch happens after September 14 at 12:00 Beijing time. The two old Flash names are already retired but temporarily aliased.

Do gateway prices match official peak/off-peak rates?

Not necessarily. Our model IDs, availability, and billing come from the live pricing page and your actual invoice. Start with a small purchase to verify the target model.

Summary

V4.1 Flash combines native vision, asymmetric compute, and a much tighter KV cache. Worth validating on real long-input and multi-turn agent workloads; most urgent for existing apps is confirming alias routing changes and the September 14 Pro switch. Practical order: confirm model ID, validate on your tasks, reconcile cache and output costs, then update config.