DeepSeek’s newest model is bigger than its predecessor and still cheaper to run. That’s not supposed to happen — until you look at where the savings actually come from.
What shipped
DeepSeek-V4.1-Flash, released this month, is a 552-billion-parameter multimodal model supporting context windows up to one million tokens (DeepSeek’s official announcement: deepseek.com). DeepSeek is already routing V4-Pro traffic to it at V4-Flash pricing while V4.1-Pro is still in development — a signal the company considers this the better model, not just a cheaper one.
The actual engineering story
The headline claim, straight from DeepSeek’s technical paper: “DeepSeek-V4.1-Flash reduces its global KV cache footprint… to 890 bytes per token, roughly 1/4 of the corresponding footprint of DeepSeek-V4-Flash,” and reduces the persistent (disk-stored) footprint to “roughly 1/8 of that of DeepSeek-V4-Flash” through a technique they call SWA Bounded Replay (arXiv paper: arXiv html/2609.19969v1). KV cache is the model’s working memory while it holds context — shrinking it by that much, at the same context length, is what actually limits how many users a model can serve at once, not a cosmetic optimization.
How they got there: the model activates only 8 billion parameters per token while reading a prompt (“prefill”) and 16 billion while generating a response (“decode”), despite 552 billion parameters total, via what DeepSeek calls a Causal Encoder-Decoder architecture — the decoder builds its working memory from a compressed summary the encoder already produced, rather than recomputing everything itself (per the model card: Hugging Face). Layered on top: Compressed Sparse Attention 2, letting neighboring layers share cached data instead of each layer keeping its own copy, plus 4-bit precision for cached values instead of the usual 16 or 32-bit.
Does it actually perform
Checking directly against DeepSeek’s own published benchmark table (Hugging Face) rather than the announcement’s framing: on Terminal-Bench 2.1, V4.1-Flash scores 90.6 against its own larger, pricier sibling V4-Pro’s 87.9. On DeepSWE issue-resolution, 74.2 versus V4-Pro’s 62.7. On CyberGym, 88.1 versus 83.3 — a Flash-tier model beating its own Pro-tier predecessor, per the same benchmark table. It is not ahead of the field generally: Claude Opus 5.0 still leads on raw reasoning (HLE: 56.3 vs. V4.1-Flash’s 36.8, same table), and on the harder Terminal-Bench 3.0/4.0 variants Opus scores 43.3 and 51.8 against V4.1-Flash’s 30.0 and 31.2. All of these numbers, including the ones favorable to DeepSeek, come from DeepSeek’s own evaluation — in several cases (noted directly on the model card) run with DeepSeek’s own harness rather than a shared neutral one, so “beats Opus” and “loses to Opus” here both carry that same asterisk equally.
Why this matters beyond one model
Shrinking the memory cost of long-context inference, rather than just adding parameters, is where a meaningful share of AI infrastructure cost actually lives once a model moves from demo to production serving thousands of concurrent long-context requests. A model that’s simultaneously bigger, ahead of its own predecessor on agentic benchmarks, and cheaper to keep in memory is a real engineering result — reported, so far, only by the team that built it. No independent reproduction of these numbers exists yet at time of writing.