DeepSeek's New Base Loses 12 of 16 Rows to the Model It Replaces

DeepSeek released V4.1-Flash on September 10. Its frontier comparison table puts a 552B model ahead of the 1.6T V4-Pro on almost every agentic benchmark, and the company announced it would start routing Pro traffic to Flash on September 14.

The base-model table on the same card tells the opposite story. And as of this writing, the English and Chinese pricing pages give two different answers about whether V4-Pro survives Sunday at all.

Everything below was read directly from the model card and both live pricing pages. The release post publishes its benchmark tables as images, so the text versions on Hugging Face are the source used here.

folio 01  what shipped

gloss: a smaller model with a bigger cache story, native vision, and a new model name you should pin.

DeepSeek-V4.1-Flash is a multimodal mixture-of-experts model with a 552B backbone, a one-million-token context window and 384K maximum output. Weights are on Hugging Face under MIT. Vision is built in rather than bolted on, which is the capability V4-Pro does not have at all.

The architecture is a Causal Encoder-Decoder: forty layers arranged as a twenty-layer causal encoder followed by a twenty-layer decoder. The decoder's global KV cache is projected from the final encoder hidden states instead of being derived from each decoder layer's own hidden states. That one choice is what lets the model activate 8B parameters per token during prefill and 16B during decode, against 49B for V4-Pro and 13B for the older V4-Flash.

Each MoE layer carries one shared expert and 384 routed experts, six firing per token. Pretraining ran on a 45T-token multimodal corpus, sparse attention trained at 64K and context extended to 1M at the 34T mark. On top of the backbone sits Engram, a conditional memory of 196B parameters reached by sparse token-based lookup.

figure a · the weight sheet

Anyone sizing storage from the press number is short by roughly two hundred billion parameters. The 552B figure is the backbone only.

Two integration details before you wire anything up. The model name is now deepseek-flash, and the legacy names still resolve but the models behind them are gone. And the release ships no Jinja chat template. You get a self-contained Python reference encoder with test cases, or a set of Rust libraries with Python bindings that handle Messages, Chat Completions and Responses formats. Inference, tool execution and transport stay on your side.

folio 02  the table nobody quoted

gloss: the new pretrained model knows less than the pretrained model it is replacing.

Every launch story this week quoted the frontier comparison table, where V4.1-Flash leads V4-Pro on Terminal-Bench, DeepSWE, CyberGym and the rest. Directly above it on the same card sits a base-model table, run in DeepSeek's internal framework with the same settings across all three models and a stated equivalence band of 0.3 points.

Sixteen benchmarks are reported for all three base models. The new 552B base leads on four. It trails the 1.6T V4-Pro base on the other twelve.

figure b · base-model ledger, widest gaps first

Green rows are the four wins. Three of them are code and math. The fourth, GSM8K at 0.4, sits just above DeepSeek's own 0.3 equivalence band. On MGSM the new base also trails the older 284B V4-Flash base, which scores 85.7.

The card explains the pattern without being asked. It states that the post-training recipe is standard supervised fine-tuning, then reinforcement learning, then on-policy distillation, with no algorithmic modifications, and that all the substantive changes sit in the data pipeline instead: large-scale automated synthesis of agent tasks and environments, scaled progressively across data, tasks and rollouts. The agentic lead is a data story layered over a base that has less in it.

folio 03  what that trade means for you

gloss: a good trade for a loop that reads files, a bad one for a model answering from memory.

If your system is retrieval-grounded, the twelve losing rows barely reach you. An agent that greps a repository, reads the file and runs the test does not need the weights to remember who wrote a 2011 paper. It needs the loop to keep its footing over hundreds of steps, which is exactly what the synthesized agent-task post-training was built for, and the agentic table is where that shows up.

If your system answers from parametric memory, the same swap costs 12.9 points of verified factual recall and 6.3 points of long-context comprehension. No amount of scaffolding recovers that. Before moving anything off V4-Pro, run your own recall set. The base table tells you roughly how big the hole should be.

folio 04  the cache is the product

gloss: 890 bytes per token is a serving win first and your win second.

The paper's own title is about pushing KV cache compression to its limit, and three mechanisms do the work. Compressed Sparse Attention 2 assigns every attention layer one of three static modes, Full, Reindex or Reuse, so layers share main KV and indexer keys and reuse top-K sparse indices. A hierarchical sparse indexer in the decoder restricts later indexing layers to a candidate pool built by the first Full-mode layer, which bounds deeper indexer cost independently of how long the context gets. And the main KV cache is held in a four-bit float format with one scale per sixteen channels.

Together those land the global cache at 890 bytes per token, about a quarter of the previous generation and, per the card's figure caption, roughly 437 times smaller than DeepSeek-V1. A separate trick called bounded replay rebuilds missing sliding-window states by replaying the most recent window of tokens rather than writing them to disk, which is where the eight-fold cut in persistent footprint comes from.

DeepSeek's launch post frames this as your saving, noting that cache-hit charges often make up a large share of agent costs. At its own published rates that only becomes true at extreme reuse. On Flash off-peak, two hundred cache-hit input tokens cost the same as one output token. On V4-Pro the ratio was about ninety to one. So the compression pushed cached input further down the bill, and the honest read is that this win lands on DeepSeek's memory and storage first and reaches you as the headline rate cut.

folio 05  the clock is part of the price

gloss: peak is the Chinese office day, so your overnight batch is the expensive one.

figure c · rate card, USD per million tokens

Peak is twice off-peak throughout. Against V4-Pro off-peak, Flash bills about one seventh the rate on cached input, roughly a quarter on uncached input and under a third on output, using the renminbi list prices the USD column is converted from.

Now the part that is easy to miss. Peak hours run 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday through Friday. The Chinese page gives the identical window in local terms: 9:00 to 12:00 and 14:00 to 18:00 Beijing time. That is the Chinese working day with the lunch break carved out as off-peak. Thirty-five hours out of a hundred and sixty-eight are peak, so 79 percent of the week bills at the lower rate by default.

For a team working US hours, the entire business day sits off-peak. Nine to five Eastern ends four hours before peak opens, nine to five Pacific an hour before it. What lands inside peak is the evening and the overnight window, precisely when unattended agent jobs, nightly evaluation sweeps and cron-driven batches tend to run. The habit of pushing heavy work to the small hours is now a two-times line item.

figure d · one agent run, four ways, illustrative arithmetic

Arithmetic on published rates, not a measured run. Read the shape rather than the totals: on the Flash off-peak row, output is 0.18 of the 0.224 and the celebrated cached input is 0.006. Budget on output.

folio 06  the headline is the best of eight

gloss: the lead over Opus-5.0 is smaller than the spread across DeepSeek's own harnesses.

The card does something most labs skip. It runs the same two agentic benchmarks through eight different scaffolds and prints the whole spread.

figure e · same model, eight harnesses

Both scaffold choices are disclosed on the card: mini-SWE to match the benchmark's official setup, Minimal mode for the terminal suite.

Read the frontier table against this one and the margins get thin. V4.1-Flash posts 90.6 on Terminal-Bench 2.1 against Opus-5.0 at 89.1, a lead of 1.5 points. The spread across DeepSeek's own eight harnesses on that benchmark is 6.5 points. Swap in Claude Code and the number drops to 88.0, below the Opus figure it was beating. Same shape as Qwen's 86.6 and GLM-5.3's 0.7-point CyberGym margin: when a published agentic lead is smaller than the harness spread underneath it, the lead is a configuration, not a capability. One thing in DeepSeek's favour, it is among the few labs that hand you the table to check this with.

folio 07  two pages, two fates for v4-pro

gloss: the same footnote number says opposite things in two languages.

The launch story every outlet ran this week was the deprecation. Starting Sunday, requests to the Pro endpoint would be answered by Flash and billed at Flash rates until a V4.1-Pro arrives. The release post still says so. Footnote two on the pricing page is where it stops being simple.

figure f · footnote (2), both editions, read sept 11

Three of the four live documents say the model goes away on Sunday. The English pricing page is the only one that says it stays, and it is the one an international team is most likely to be reading.

The practical response is the same either way. Do not leave a hard-coded Pro endpoint in production betting on one outcome, and log the model field on every response over the weekend. The difference between the two policies is not a capability question, it is a silent substitution: the call succeeds, the bill drops, and the thing that answered has 12.9 fewer points of factual recall than the thing you asked for. One note for whoever owns the budget: the two pages imply slightly different renminbi to USD conversion rates across the two models, so treat the USD column as a rounded conversion of a renminbi list price.

folio 08  honest notes

gloss: what cuts against the reading above.

• Four of the twelve losing base rows sit inside 1.4 points, and DeepSeek states a 0.3 equivalence band. The pattern holds, but the count of twelve is doing more work than any single row except SimpleQA-Verified.
• The claim that third parties measured Flash ahead of Pro is unattributed on every page carrying it. No names, no benchmark set, no harness. Every number in this issue is DeepSeek's own.
• On the newest terminal benchmarks the frontier gap is still wide. Flash scores 30.0 and 31.2 on Terminal-Bench 3.0 and 4.0 against Opus-5.0 at 43.3 and 51.8. Its reported HLE score also sits below V4-Pro's.
• All instruct results use maximum reasoning effort on a dial that runs from 1 to 100. Nothing on the card shows what the curve looks like in between, or what the same tasks cost at lower settings.
• Press reporting this week ties the release timing to DeepSeek's preparation for a Shanghai STAR Market listing. That is reporting about the company, not a fact about the model, and nothing above rests on it.

folio 09  five things to do this week

1. Grep the codebase for the Pro model string. Pin an explicit name and start logging the model field on responses before Sunday, so a substitution shows up in your logs rather than in your evaluations.
2. Move flexible batches out of 01:00 to 10:00 UTC on weekdays. Doubling is not a rounding error, and the window you are moving out of is the one your overnight jobs already sit in.
3. Budget on output tokens. Cached input stays negligible until you are reusing two hundred cached tokens for every token generated.
4. If any part of the product answers from the model's own knowledge rather than from retrieved context, run a recall set before switching. The base table says to expect a real drop there.
5. If you plan to self-host, size from the 763B figure on the repository, not the 552B backbone number in the headlines.

next folio

The reasoning effort dial runs 1 to 100 and every published number uses 100. Next issue we run one identical forty-turn coding task at effort 20 and at effort 100, then print both invoices beside both outcomes, so the dial gets a price instead of a description.

sources read for this issue

Release post, DeepSeek API docs · Model card, Hugging Face · Technical report · Pricing, English · Pricing, Chinese · Company announcement · Context caching guide · Thinking mode guide · Rate limits and concurrency · deepseek-recipe prompt encoder · Prompt encoding reference · SiliconANGLE · TechNode · Technology.org on the listing context