Playing this video loads it from YouTube (Google), which may store cookies on your device and receive your IP address. Click to play.
Video: "Deepseek v4.1 is SCARY GOOD!" by Julian Goldie on YouTube.
What makes V4.1 Flash different from V4 Pro
On paper, Flash looks like the smaller, cheaper option. The total parameter count is 552 billion, but only 8 billion activate per input token (the model uses 16 billion on output tokens). That mixture-of-experts approach means the inference cost is a fraction of what a comparably-sized dense model would require. The upshot: $0.30 per million input tokens and $1.20 per million output tokens - well below V4 Pro's pricing.
In practice, Flash does not behave like a smaller model. The new Causal Encoder-Decoder architecture gives it native multimodal reasoning - it processes images and text through the same mechanism rather than treating vision as a bolt-on. The context window runs to one million tokens. On the standard reasoning and coding benchmarks that V4 Pro previously led, Flash matches or exceeds it across most categories. The cleaner inference cost profile is an advantage rather than a tradeoff.
What Julian Goldie actually tested it on
Goldie's test focused on practical coding and generation tasks rather than abstract benchmarks. He built a working application from a text prompt in a short session. He generated a 3D website - complete with visual structure and navigation - in a few minutes. He ran it through a series of coding challenges that had tripped up earlier models.
The results were consistent with the benchmark scores but more useful to watch: the model produced functional, structured output quickly and made fewer errors requiring correction than the previous generation. For teams that use AI to accelerate development work, the key observation is that Flash produces working first drafts more reliably, which reduces the time spent on debugging generated code rather than building on it.
What the MIT licence actually changes
Most frontier AI models are closed: you access them through a paid API with no option to run the weights yourself. DeepSeek's MIT licence on V4.1 Flash means the weights can be downloaded and run locally, modified, and deployed without restriction - including commercially.
For a business with the right hardware, running V4.1 Flash locally eliminates API costs entirely. On a capable server or a modest cluster, you get the same model performance with no per-token charge. The hardware investment required is real - running a 552B MoE model is not trivial - but for high-volume workloads, the economics shift considerably. More immediately, the open weights mean that fine-tuning on proprietary data is possible without sending that data to a third-party API, which matters for organisations handling sensitive information.
Where the model fits into a development workflow
The combination of low cost, strong reasoning, and native multimodal input makes V4.1 Flash useful at several points in a typical development cycle. It handles code generation and review tasks well. The long context window means it can work with large codebases - entire repositories passed in a single request rather than chunked. The visual understanding means it can take a screenshot or design mockup and generate matching code, which collapses the gap between design and implementation for smaller teams.
There are limits worth noting. The model is strong on general programming and reasoning, but specialist domains - complex legal analysis, highly regulated financial modelling - will still warrant more targeted tools or human review. And while the MIT licence permits local deployment, the infrastructure cost of running the full model is non-trivial. For most teams, the API will be the practical route, and Token Harbour's free tier (covered in our companion article on the Hermes Agent setup) provides access without committing to a paid plan.
Where this connects to NordSys
Models like V4.1 Flash change what is practical to automate: lower cost per token means more tasks make economic sense to run through AI, and better multimodal reasoning means more kinds of input can feed into automated workflows. Our Programming service works with models and tools like these to build custom automations - data pipelines, internal tools, integrations between your existing systems - that fit the actual scale and budget of your operation rather than enterprise pricing assumptions.
See our Programming service →