DeepSeek V4.1 Flash: the new version sees images and beats V4 Pro
DeepSeek has released V4.1 Flash — a model that replaced two earlier ones at once: the regular V4 Flash and an experimental version with image recognition. Now it is a single model that both writes code and reads images: it beats V4 Pro on agentic benchmarks, holds a one-million-token context, and uses roughly four times less KV cache. On NeuralSpace, V4.1 Flash is already available in the chat, in the "Code" section, and via the API — here is what changed compared with V4 Flash.
What this model is
Strictly speaking, this is not a revision of the previous Flash but the first model of a new architecture family. V4.1 Flash has 552 billion parameters against 284 billion in V4 Flash, so the model is noticeably bigger. Yet it activates fewer parameters per step: about 8 billion while reading the input and 16 billion while generating the answer.
The new scheme is called Causal Encoder-Decoder: the 40 transformer layers split into 20 encoder layers and 20 decoder layers, and the shared key-value cache is built once — from the encoder's hidden states — instead of being recomputed in every layer. That is why input processing is cheap. The asymmetry is deliberate: an agent reads a lot — files, conversation history, instructions — and writes comparatively little — edits, commands, decisions. The expensive half was made light, so agentic workloads got cheaper too.
The context is one million tokens. Reasoning effort is continuously adjustable from 1 to 100: a simple question can run cheaply, while a hard chain can run at maximum without switching models.
It sees images now
This is the most visible difference from the previous version. V4 Flash had a text-only input: it could not look at images, and DeepSeek shipped a separate experimental vision model for that. In V4.1 Flash, vision is built into the model itself — it natively accepts text and images and replies with text.
In practice, you can hand the model a UI screenshot, a chart, a photo, or a document page as is. It will describe the image, extract text from a screenshot, or read a diagram. That is especially handy for agents: one model reads both the code and the UI screenshot, with no need to switch between two services.
DeepSeek retired the old model names: requests to v4-flash and v4-flash-vision-exp are now routed to V4.1 Flash, so existing integrations will notice nothing.
It beats V4 Pro
The loudest part of the release: V4.1 Flash outperformed the larger V4 Pro on agentic tasks. Here are the numbers from DeepSeek's official evaluations:
- Terminal-Bench 2.1 (working in a terminal) — 90.6 against 87.9 for V4 Pro and 82.7 for the previous V4 Flash;
- CyberGym (cybersecurity) — 88.1 against 83.3 for V4 Pro;
- DeepSWE v1.1 — 74.2 against 54.4 for V4 Flash;
- Terminal-Bench 4.0 — 31.2 against 7.0 for V4 Flash.
On top of that: 90.9 on GPQA Diamond, a Codeforces rating of 3471, and 65.6 on MathArena Apex. The numbers are not independent — DeepSeek reported them itself, so treat them as a claim rather than a verdict. But the scale of the jump over the previous Flash is visible without third-party runs.
DeepSeek also announced an orderly phase-out of V4 Pro: from September 14 its API routes requests for V4 Pro to V4.1 Flash and bills them at Flash pricing. The practical takeaway is simple: Flash is no longer the "junior" model. It now carries the main load, and the regular and vision versions no longer exist separately.
Smaller cache, cheaper agentic tasks
For agentic scenarios, the main saving is not in the token price but in the cache. An agent re-reads the conversation history, instructions, and project files again and again, and it is the cached input that eats most of the bill. V4.1 Flash's global KV cache takes about 890 bytes per token — roughly four times less than V4 Flash — and its persistent cache is compressed about eightfold.
The model price dropped against the previous Flash as well: cached input is 60% cheaper, uncached input about a third cheaper, and output 11% cheaper. The gain is most visible in workloads that repeatedly process a large context; where the length of new output matters more, the saving is more modest.
V4 Flash versus V4.1 Flash: what changed
| Parameter | V4 Flash | V4.1 Flash |
|---|---|---|
| Parameters | 284B | 552B |
| Activated per step | 13B | 8B input / 16B output |
| Architecture | MoE decoder | Causal Encoder-Decoder |
| Understands images | No, separate vision model | Yes, natively |
| Context | 1M tokens | 1M tokens |
| Terminal-Bench 2.1 | 82.7 | 90.6 |
| DeepSWE v1.1 | 54.4 | 74.2 |
| KV cache per token | ~4× larger | 890 bytes |
Price on NeuralSpace and how to try it
On NeuralSpace the model runs at the same rate as the previous V4 Flash: $0.14 per million input tokens and $0.28 per million output tokens. You can start with as few as 3 tokens on your balance; usage is charged from the shared balance, with no separate subscription and no foreign card. To try it, one step is enough:
- Chat: the model page — attach an image or screenshot and the model will read it;
- Code: the "Code" section — the model connects to your project and works with the repository, files, and terminal;
- API: get a key on the API keys page; the format is OpenAI-compatible, and the docs live at neuralspace.pro/en/v1/docs.
Frequently asked questions
Is V4.1 Flash a new model or an update? It is a model of a new architecture family rather than a revision of the previous Flash: the architecture changed, the backbone grew, and vision was added.
Does it see images? Yes, natively: send a photo, screenshot, chart, or document. The previous V4 Flash was text-only.
What happened to V4 Pro? DeepSeek is phasing it out and routing requests to V4.1 Flash, which beat V4 Pro on agentic benchmarks.
How much does it cost to try? $0.14/$0.28 per million tokens, starting from 3 tokens on your balance — on the chat model page.
Will old v4-flash requests break? No: calls to v4-flash and v4-flash-vision-exp are routed to V4.1 Flash and billed at its rate.