DeepSeek V4.1 Flash: A 552-Billion Parameter Model That Fits in a Budget

DeepSeek dropped a quiet bombshell this week. On September 10th, the Chinese AI lab released DeepSeek-V4.1-Flash — a 552-billion parameter mixture-of-experts model that only activates 8 billion of those parameters per token. The result is a model that punches far above its weight class, and it's fully open-weight under the MIT license.

Small Footprint, Giant Performance

The "Flash" name is a bit misleading if you think size. V4.1-Flash has a 552B parameter backbone — enormous on paper — but thanks to its sparse mixture-of-experts architecture, only 8B parameters fire per input token (16B during generation). This keeps compute costs and inference latency low while retaining the knowledge of a much larger model.

DeepSeek also aggressively optimized the KV cache, storing entries in a 4-bit floating-point format. The result is a global KV cache footprint of just 890 bytes per token — roughly a quarter of what the previous V4-Flash needed. For operators running large fleets of inference servers, that's a meaningful reduction in VRAM pressure.

The model also supports a 1 million token context window and handles both text and image inputs natively, trained from scratch on 45 trillion tokens of mixed text and image data.

Benchmark Numbers Worth Taking Seriously

According to BenchLM's independent evaluation, V4.1-Flash scored 90.6 on Terminal-Bench 2.1, narrowly beating Anthropic's Opus 5 (89.1) and OpenAI's GPT-5.6 Sol (88.8). On DeepSWE v1.1 — a software engineering benchmark that tests real-world code tasks — it resolved 74.2% of issues, compared to 74% for Opus 5 and 62.7% for DeepSeek's own V4-Pro.

In other words: V4.1-Flash outperforms DeepSeek's own flagship on several key benchmarks, at a fraction of the cost.

Pricing and Availability

The API is live now under the model string deepseek-flash, priced at $0.15 per million input tokens (off-peak) and $0.60 per million output tokens. That puts it in the same price tier as mid-range models from Western providers, while competing with their top-tier offerings.

The weights are published on Hugging Face under the MIT license, which means teams with privacy requirements or data sovereignty concerns can self-host the entire model stack without touching DeepSeek's infrastructure.

Why This Matters

Two trends converge here. First, the continued collapse of the cost-performance curve in frontier AI: a model that was competitive with the absolute best six months ago now runs at commodity pricing. Second, the open-weight movement is increasingly producing models that match or exceed closed competitors — and this time the license is MIT, not the more restrictive terms that accompanied some prior DeepSeek releases.

For developers building agentic pipelines or running long-context document work, V4.1-Flash's 1M token window and sub-$0.20 input pricing is a compelling combination. The next question is how the Western labs respond — and how fast.