DeepSeek V4.1 Flash activates 8 billion parameters per query and leads in coding

DeepSeek announced on Thursday (10th) the V4.1 Flash, the smallest model in its new series, claiming it outperforms the company's former flagship in coding, cybersecurity and autonomous agent tasks, while cutting inference cost and increasing speed. It's another move in the price and performance war dominating the Chinese AI market, according to the South China Morning Post.

The trick is in the MoE, not in raw size

V4.1 Flash is built on a 552-billion-parameter framework, but uses a Mixture-of-Experts (MoE) design with a new architecture DeepSeek calls "Causal-Encoder-Decoder."

The practical difference matters for whoever pays the compute bill. In a traditional (dense) model, every query passes through the entire system. In a MoE, routing sends the task only to the most suitable subnetworks. In Flash's case, the company claims that only 8 billion parameters are activated to process the input and 16 billion to generate the response, which significantly reduces the computational power spent per request.

In practice, it's the same principle that made DeepSeek stand out before: delivering cutting-edge reasoning at an operational cost of a fraction of a cent per request. The model also brings native multimodal visual understanding, according to the company.

The numbers DeepSeek put on the table

The highlight of the announcement is Terminal-Bench 2.1, a benchmark that evaluates AI on real computing tasks, closer to a dev's actual work than multiple-choice questions. DeepSeek reported the following results table:

| Model | Terminal-Bench 2.1 | |---|---| | DeepSeek V4.1 Flash | 90.6 | | OpenAI GPT-5.6 Sol | 88.8 | | Moonshot AI Kimi K3 | 88.3 | | DeepSeek V4 Pro | 87.9 |

The usual caveat applies: these numbers come from DeepSeek itself, not from an independent evaluation. The company claims Flash outperformed V4 Pro (its larger, more expensive model) on coding, cybersecurity and agent task tests. What remains open is how these results hold up outside the lab and in third-party evaluations.

Why this shifts the cost baseline

The backdrop is the compute squeeze: with hardware costs rising and chip export restrictions tightening access for Chinese companies, the path they've found is efficiency. Instead of scaling active parameters, the game has become "how little compute can I spend while keeping cutting-edge reasoning." The MoE in V4.1 Flash is the materialization of that: 552B on paper, 8B to 16B active in practice.

This shift is what makes the news relevant for those building software outside the OpenAI/Anthropic axis. When a model with competitive agent numbers activates just a few billion parameters per request, it moves the bar on what's "too expensive to run." That applies both to API cost and to the viability of serving open models on your own infrastructure (the DeepSeek line has historically had open weights, though this specific announcement doesn't detail V4.1 Flash's licensing).

What changes for the Brazilian developer

The practical takeaway is that the Chinese efficiency race is pushing down the cost floor for tasks that matter to code writers: code generation and review, automation of agents that operate terminals, and security workflows. If V4.1 Flash delivers agent performance close to GPT-5.6 Sol while spending less compute per call, the calculation of "is it worth depending on a single foreign vendor?" changes.

For the BR engineering team, the concrete points to watch:

• Cost per request in reais: with exchange rates weighing in, fraction-of-a-cent-per-call models open room for automations that previously didn't add up.
• Terminal-Bench as a signal: it's a real terminal task benchmark, so it's a better proxy for dev agents than academic tests. If the numbers hold up in independent evaluation, it's a candidate to test in agent pipelines.
• Strategic dependency: having viable alternatives to OpenAI and Anthropic reduces the risk of price lock-in and access policy lock-in.

What remains pending is the usual with vendor announcements: where the weights sit, what API access looks like outside China, real latency for serving from here, and, above all, whether the 90.6 on Terminal-Bench holds up when it isn't DeepSeek itself running the test. Until then, the central fact is clear: the efficiency fight keeps pushing AI costs down, and that's good for those who build.