DeepSeek v4.1 Flash pricing, benchmarks, and 890 bytes per token

DeepSeek calls V4.1-Flash the smallest model in its new architecture family, designed for greater capability, faster inference, higher throughput and scaling to larger models. A separate Hacker News submission carried the release under the title "DeepSeek launching v4.1 flash cheaper and more capable than v4 pro", and that submission title is the only place in these sources where the v4 Pro price comparison appears at all. Weights are on Hugging Face as deepseek-ai/DeepSeek-V4.1-Flash. No source names the licence attached to them.

DeepSeek describes the design as asymmetric: a 552B-parameter MoE where input costs 8B active parameters and output costs 16B. Commenters quoting the model card fill in the shape: a 40-layer Transformer organized as a 20-layer causal encoder followed by a 20-layer decoder, in which the decoder's global KV cache is projected from the final encoder hidden states rather than derived from each decoder layer's own hidden states. The card ties that projection directly to "substantially improving cost efficiency for input-heavy agentic workloads", which is the shape of a tool-calling loop replaying a growing transcript.

The full parameter accounting, again per the model card as quoted in the thread, is 552B backbone parameters plus 196B Engram parameters, "activating 8B parameters per token during prefill and 16B during decode". Unsloth AI, replying to the announcement, singled out the 196B Engram split as the thing making the model more accessible. DeepSeek's own thread does not mention the Engram figure at all.

Long-horizon agent runs are dominated by re-reading context, so per-session KV cache is what caps how many concurrent sessions fit in a fixed HBM budget. DeepSeek makes the economic link explicit rather than leaving it as an inference.

At 890 bytes per token, one commenter put a full million-token context at roughly 1GB of global KV cache. Cutting per-session cache to a quarter and the SSD tier to an eighth changes the capacity arithmetic for anyone sizing a node for concurrent agent sessions, though none of these sources report measured throughput or latency under concurrency. The prefill and decode asymmetry pushes the same direction: the half of the workload that grows with context length runs on 8B active parameters rather than 16B.

Two pricing facts are on the record from DeepSeek. First, peak and off-peak pricing continues, off-peak rates are 50% of peak rates, and the company recommends scheduling flexible workloads off-peak. Second, it says the more efficient architecture lets it serve more users at lower cost and that it is passing the savings on. The retrieved thread cuts off before the per-token table, so the absolute cost per token is not something these sources establish; one commenter characterized the new price as almost back to the old V4-Flash price while claiming the model now rivals competitors that cost ten to thirty times more, which is a practitioner's read rather than a published figure.

The benchmark side is thinner. DeepSeek says new pre-training methods and larger-scale RL post-training deliver "benchmark results ahead of" something the retrieved text truncates away, and the API post begins "Tests by" before truncating as well. No benchmark name, score or harness appears anywhere in this source material. One commenter points to a figure on page 5 of a file in the Hugging Face repo; a different commenter claims the model goes past Kimi K3 and GLM 5.3 in most tests at 250 to 400 t/s and is 2x to 2.5x faster than the older Flash. Treat both as unverified until the tech report is read directly.

The loudest objection in the Hacker News thread is size. One commenter notes the original V4-Flash was 284B against this model's 552B, argues it is not really flash anymore, and says the jump in benchmark scores follows straight from the extra parameters. Another adds the quantization angle: V4-Flash shipped mostly FP4 at around 160GB, while this release appears to be FP8 at around 510GB, marked with a question mark in the comment itself, turning a model that fit a dual Spark or Strix Halo box into something needing four machines or third-party quants. The same commenter then works out that most of the added size is Engram parameters, which can stay on SSD, so the resident non-Engram portion may be closer in footprint than the raw number suggests.

The counter-argument in the thread is that flash has always meant speed, not parameter count, and that this model uses fewer active parameters, "8B or 14B instead of always 13B" in one commenter's phrasing, against the card's stated 8B prefill and 16B decode. One commenter expects SSD streaming to work well here because the MoE expert sparsity looks similar to DeepSeek V4 Pro with far fewer total and activated parameters, and because Engram parameters can reside on disk as they do for Qwen Flash-Next. Another offered a concrete data point from Qwen 3.8 Flash Next on a Mac Studio: 300 prompt processing and 33 token generation with SSD offload, versus 550 and 40 with everything in RAM, and said they would be happy to see 300 prompt processing on the DeepSeek model.

There is no tool-calling benchmark for V4.1-Flash in these sources. What exists is behavioural reporting from the thread: one commenter runs it for automated bug triage, collecting every unique error message overnight and asking the model to find the bug, and calls it great for that work. One commenter says it seems more willing to just do whatever you ask, with a ROM download for a game they own as the test; a reply adds that reviewing a codebase containing an out-of-bounds access bug gets blocked by US models and that "find vulnerabilities in..." prompts are out of the question, while Chinese models oblige. Another commenter expects those gaps to close in a few months or years. None of this is a measurement, and refusal posture is not a substitute for a function-calling eval.

The one third-party number available touches the predecessor, not this release. A reasoning-prefill experiment over 45 problems, 15 STEM, 15 non-STEM and 15 synthetic puzzles, inserted the first 1% of GPT-5.5 Pro's reasoning into each target model's reasoning channel and scored how much of the teacher's visible answer appeared in the first 100 tokens, as the mean of unigram, bigram and trigram source recall. DeepSeek V4 Flash scored 27.30% unprefilled and 26.13% prefilled, a delta of -1.17 percentage points, the only negative delta in the table; Qwen3.8 A95B moved +18.18 points, Kimi K3 +4.54, Inkling +0.46. That says something about how independent V4 Flash's reasoning trace is, and nothing at all about V4.1-Flash.

The API change is immediate. Set the model to deepseek-flash for native multimodal support. V4-Flash and V4-Flash-Vision-Exp are retired, with deepseek-v4-flash and deepseek-v4-flash-vision-exp temporarily routing to V4.1-Flash for compatibility, so anything pinned to those two strings is already running on new weights and behavioural regressions in an existing agent loop should be attributed there first. The reasoning effort control is a continuously variable integer from 1 to 100 that trades inference cost for accuracy.

Open questions these sources do not answer: the exact per-token peak and off-peak prices, the licence on the Hugging Face weights, the benchmark table and what it was compared against, the released precision, and any measured tool-calling or long-context retrieval score. DeepSeek says it will work with the open-source community on V4.1-Flash inference support and explore more deployment options, and invited anyone planning a 2,000-GPU deployment with a storage cluster to get in touch. For teams sizing hardware, the number to verify first is whether 890 bytes per token holds under real concurrency, because that, not the 552B headline, determines how many agent sessions fit on a node.