DeepSeek stopped waiting for a memory-cost breakthrough and shipped one. On September 10, 2026 the Hangzhou lab released V4.1-Flash, an open-weights multimodal Mixture-of-Experts model with 552 billion backbone parameters, an eight-times smaller KV cache than the July V4-Flash, and pricing that starts at $0.15 per million input tokens.
552B parameters, 8B active, 1M context
V4.1-Flash keeps DeepSeek's now-familiar MoE shape but tunes it for agent workloads: 552B backbone weights, only 8B active per token during prefill and 16B during decoding, and a 1M-token context window with up to 384K tokens of output. Text and images are handled natively rather than through a separate vision head.
KV cache slashed to 890 bytes per token
The real headline is memory. DeepSeek says the model compresses its global key-value cache to roughly 890 bytes per token — about a quarter of what V4-Flash needed and 437 times smaller than DeepSeek V1 — using FP4 quantization and a cross-layer attention reuse pattern the company calls pure CSA2. In practical terms V4.1-Flash needs roughly one quarter of the HBM and one eighth of the SSD storage of its predecessor to hold the same conversation state, which is what makes agentic workloads with million-token histories affordable on a single node.
Beats Opus 5 and GPT-5.6 Sol on DeepSWE
On the DeepSWE agentic-coding benchmark V4.1-Flash narrowly edges Anthropic Claude Opus 5 and OpenAI's GPT-6 Astra-family GPT-5.6 Sol, an unusual result for a Chinese open-weights model at a fraction of the price. DeepSeek is also retiring V4-Pro from public serving and steering developers to V4.1-Flash as the new default, mirroring an earlier V4-Pro price hike that pushed users toward the leaner Flash line.
MIT license, $0.15 per million input tokens
V4.1-Flash ships under the MIT license and lands at $0.15 per million input tokens off-peak on DeepSeek's API, with peak-hour multipliers and native Codex plus OpenAI Responses API compatibility carried over from the summer release. The weights are on Hugging Face and the API is live at the standard DeepSeek endpoint.
Reporting based on coverage from DeepSeek, MarkTechPost, Cryptobriefing and LLM Stats.