Aug 05, 2026
DeepSeek V4 Flash, explained: what the 1M-context model does and where your prompts go
What DeepSeek V4 Flash is, the hybrid attention behind its million-token context, and where your prompts actually go when you use it from Europe.
DeepSeek V4 Flash is the efficiency model of DeepSeek's V4 generation: a mixture-of-experts language model with 284 billion total parameters, 13 billion active per token, a context window of one million tokens, and open weights under the MIT license. DeepSeek released it in preview on April 24, 2026 and shipped the official build on July 31, 2026. In ilisai it is a selectable chat model, billed per token in credits.
This guide covers what the V4 generation changed, the architecture that makes a million-token window affordable, what the model can and cannot do, how it shows up in ilisai, and — since the lab behind it is Chinese and ilisai's users are mostly European — where your prompts actually go.
What is DeepSeek V4 Flash?
DeepSeek V4 Flash (model id deepseek-v4-flash) is the smaller of the two models DeepSeek announced on April 24, 2026, alongside the flagship V4-Pro. Both are open weights under MIT. Pro carries 1.6 trillion total parameters with 49 billion active; Flash gets by with 284 billion total and 13 billion active, which is what makes it fast and cheap to serve. On DeepSeek's own API, Flash costs $0.14 per million input tokens and $0.28 per million output tokens — about a third of Pro's $0.87 output rate (DeepSeek pricing documentation).
The timeline:
- April 24, 2026 — DeepSeek releases the V4 generation in preview: V4-Pro and V4-Flash, open weights, with a 1M-token context as the default across its official services.
- April 26, 2026 — the technical report lands on arXiv, describing the hybrid attention design covered below.
- July 31, 2026 — DeepSeek ships DeepSeek-V4-Flash-0731, the official build that closes the preview phase for Flash (reported by Caixin Global); the
deepseek-v4-flashAPI endpoint now points to this build.
One footnote for direct API users: DeepSeek's pricing documentation says the first-party API will soon adopt peak and off-peak pricing — double the rate during Beijing peak hours — with the effective date still to be announced. ilisai prices are unaffected.
How a million tokens became affordable
Attention cost is why long context has stayed expensive: in a classic transformer, every new token attends to every earlier one, so compute grows with the square of the sequence and the KV cache keeps growing with its length. The V4 technical report (arXiv:2606.19348) attacks this with two attention mechanisms used in alternating layers:
- CSA (Compressed Sparse Attention) compresses groups of key-value entries into single blocks, then lets each query attend only to the most relevant blocks — selected by a lightweight "lightning indexer" — plus a sliding window of recent tokens for local detail.
- HCA (Heavily Compressed Attention) compresses far more aggressively and drops selection entirely: every query attends densely to the whole compressed stream, keeping a cheap global view of the full context.
A third piece, mHC (manifold-constrained hyper-connections), reworks the residual connections between layers to keep signal propagation stable at this scale. The payoff is concrete: per the technical report, V4-Pro at a 1M-token context needs only 27% of the single-token inference compute and 10% of the KV-cache memory that DeepSeek V3.2 needed — and Flash is the smaller, cheaper sibling of that same design. In practice, you can hand this model an entire book, codebase export, or folder of reports as a matter of routine.
What the model can do
DeepSeek V4 Flash reads up to a million tokens of context, reasons by default, calls tools, caches repeated context automatically, and accepts text only. The detail, per DeepSeek's documentation, the Fireworks model page, and ilisai's own integration:
| Capability | DeepSeek V4 Flash |
|---|---|
| Context window | 1M tokens |
| Max output | 32,768 tokens per response in ilisai — the product-wide chat output cap; the Fireworks endpoint itself allows up to 131,072 |
| Inputs | Text only — no image or file input |
| Reasoning | Thinking mode, on by default, with selectable effort levels |
| Tool use | Function calling, including multi-turn reasoning + tool calls |
| Caching | Context caching enabled by default, no code changes needed |
| License | MIT, open weights |
Thinking mode. The model reasons before answering, and on DeepSeek's API this is the default (default effort: high). V4 Flash accepts three effort levels — low, high, and max. One quirk worth knowing: in thinking mode the API ignores temperature, top_p, and the penalty parameters entirely, so the prompt is the steering wheel (DeepSeek thinking-mode guide).
Tool use. The model can run several rounds of reasoning and tool calls before it commits to an answer (same guide). Add the long context and you get the profile of a model built for agent workloads, not just chat.
Implicit caching. DeepSeek enables context caching for every request with no code changes: repeated prompt prefixes are served from cache and billed at a fraction of the normal input rate (DeepSeek context-caching documentation). Fireworks applies the same idea on its platform — cached input is billed at $0.028 per million tokens against $0.14 for fresh input (Fireworks model page).
Text only. The model takes no image or file input — the Fireworks model page states image input is not supported. In ilisai the attachment button is disabled when DeepSeek V4 Flash is selected, so you find out before building a workflow around it, not after. Paste text instead — with a million tokens of room, pasting is rarely the bottleneck.
DeepSeek V4 Flash in ilisai
DeepSeek V4 Flash is available in ilisai's chat: sign in, open the model selector, pick it, and write. No DeepSeek account, no API key, no per-provider billing — the credits included in your plan spend on this model exactly as they do on any other in the catalog.
For very long documents it is a particularly comfortable model: in ilisai a single conversation carries roughly 770,000 tokens of input, after the reserve ilisai holds back for the response — a context that still holds several novels. The developer alternative is Fireworks' own API, where the same model costs $0.14 input / $0.28 output per million tokens, with its own API key and a separate invoice. In ilisai none of that is needed: every plan, Free included, comes with credits that spend on this model, and jobs beyond your included credits draw from your wallet.
If you want to see the model at work before writing your own prompts, our DeepSeek prompt library collects eight copy-paste prompts tuned to what this model does best, long-document work included.
Using it from Europe: where your prompts go
In ilisai, prompts sent to DeepSeek V4 Flash go to Fireworks AI, a US-based inference provider that keeps zero data retention by default — never to DeepSeek. For a Chinese lab's model that question deserves a precise answer, so let's take it piece by piece:
- Your prompts go to Fireworks, not to DeepSeek. ilisai serves the open-weights model through a direct, pinned integration with Fireworks AI, a US-based inference provider, on its serverless platform. Requests are never routed through DeepSeek's API or any Chinese provider — the MIT license is what makes this possible: anyone can host the weights, and ilisai chose who does.
- Fireworks keeps zero data retention by default. Per Fireworks' data-handling documentation, prompt and generation data for serverless models exist only in volatile memory for the duration of the request and are not logged to persistent storage; with prompt caching active, cached prefixes can persist in volatile memory for a few minutes.
- The serving region is Fireworks' call. Fireworks serves DeepSeek V4 Flash on its global serverless platform, with no region commitment for this model. ilisai's product — accounts, chats, billing — operates under GDPR as an EU company; inference for this model runs under the Fireworks terms linked above.
ilisai serves frontier and cost-efficient models from several providers under one account with euro pricing, and facts like these get told per model before you commit a workflow to one. Our comparison of European AI options covers the wider sovereignty picture.
Start using AI today
Create your free account and access multiple AI models from a single interface.
Frequently asked questions
What is DeepSeek V4 Flash?
The efficiency model of DeepSeek's V4 generation: a 284B-parameter mixture-of-experts model with 13B active parameters, a 1M-token context window, thinking mode, and tool use, released as open weights under MIT. The preview shipped on April 24, 2026; the official build followed on July 31, 2026.
Do my prompts go to China?
No. ilisai runs the open-weights model through Fireworks AI, a US-based inference provider, and requests never touch DeepSeek's API or any Chinese provider. Fireworks keeps zero data retention by default on its serverless platform, which is global and does not guarantee a specific region.
Can I attach images or files?
Not with this model: DeepSeek V4 Flash is text-only, and ilisai disables the attachment button when it is selected. Paste your text directly into the chat; the 1M-token context leaves plenty of room. For work that needs image input, pick another model from the catalog.
How much text fits in the context window?
One million tokens on the model's spec sheet: several long novels' worth of text, or a large project's documentation. In ilisai, a single conversation carries roughly 770,000 tokens of input after the reserve held back for the response, and a single response is capped at 32,768 tokens — ilisai's product-wide chat limit. The model supports up to 131,072 output tokens when you call Fireworks directly.
Can I use it on the Free plan?
Yes. The Free plan includes 400 credits a month, usable up to 150 a day, on any model in the catalog, DeepSeek V4 Flash included; larger jobs can draw from your wallet on any plan.
What about DeepSeek V4 Pro?
V4-Pro is the 1.6-trillion-parameter flagship of the same generation, also MIT open weights. ilisai does not serve it; Flash is the catalog's DeepSeek entry, chosen for its price-to-capability ratio.