DeepSeek Replaced V4.1 Flash Under the Same Name. It Is Now the Second-Best Big Open Model, and Your Model ID Did Not Protect You.

DeepSeek changed what deepseek-flash means on October 17, 2026. The API changelog entry is one paragraph: "DeepSeek V4.1 Flash has been upgraded to DeepSeek-V4.1-Flash-1017. The official web and APP models have also been updated." Then the sentence that matters to anyone building on it: "The API calling method and pricing remain unchanged, simply use deepseek-flash to access the latest model."

That sentence is true, and it is doing a lot of work. "Simply use deepseek-flash" means you are already on the new model, whether you asked for it or not. The two retired names, deepseek-v4-flash and deepseek-v4-flash-vision-exp, moved with it, for the second time in five weeks. DeepSeek's docs list no dated Flash ID, so there is no documented way to stay on the September build through DeepSeek's API. Same name, new weights, overnight.

This time the swap is good news on the independent numbers. Artificial Analysis rescored the model a day later at 45 on its Intelligence Index, up from 39 for the September build, which moves it from tenth to a tie for second of 121 large open-weight models with GLM-5.3 at 45, and its cost per index task fell from $0.49 to $0.26 at an unchanged list price. That is exactly why this is the moment to write down what "same name, new model" means for your evals, your prompts, and your bill, because the next swap may not be an upgrade. We read the changelog, the pricing page, the Hugging Face card and config, the V4.1 tech report, and Artificial Analysis's pages for both builds. Here is what held up.

What You Are Calling

The alias is the only name. DeepSeek's quick start and pricing pages document two model names, deepseek-flash and deepseek-v4-pro. The pricing page's footnote on the retired names says their "requests are served by the DeepSeek-V4.1-Flash-1017 model and billed at the Flash price." On September 10 those same names were "temporarily routed" to the September V4.1 Flash. Moved twice, pinned never.

The Hugging Face card calls it "a minor update to DeepSeek-V4.1-Flash." Both repos report the identical safetensors parameter total, 763,205,315,794, so the architecture and size did not change. Our reading is that this is a re-post-trained model rather than a resized one; DeepSeek does not say so in those words.

Self-hosting is the only pin. The September weights are still a separate repo. If your product depends on September's behavior, that checkpoint, on your own hardware or on a host that confirms it still serves that exact build, is the only way to keep it. DeepSeek recommends DeepSeek Harness 0.2.1-rc.1 or later "for the best experience with the new model," and says it optimized for DeepSeek Harness, Tencent CodeBuddy, and WorkBuddy. There is no Jinja chat template; prompt encoding ships as a Python reference and the deepseek-recipe Rust libraries.

The Benchmarks, Sorted by Who Reported Them

House rules: independent numbers first, stated as fact with the index version; DeepSeek's numbers labeled as DeepSeek's, with the harness, because the vendor table mixes them.

Artificial Analysis, Intelligence Index v4.3.2, October 18. AA added DeepSeek V4.1 Flash 1017 (max) a day after the swap and benchmarked it on DeepSeek's API plus four third-party hosts (Baseten, Fireworks, Parasail, DeepInfra). The September build's page is still up, which makes this the rare case where one model name carries two independent scorecards.

Six points on the index for the same list price, and the per-task bill cut almost in half: $0.26 against $0.49, a 47% drop, while output tokens fell 16%, from 250M to 210M. Something beyond output length moved, and neither AA nor DeepSeek says what. AA still ranks 1017 "very verbose," 39th of 47 on that measure, and 210M is 50% above the class median of 140M. Cheaper per task is not the same as terse.

Now the vendor table. These are DeepSeek's self-reported numbers from the October 17 changelog and the Hugging Face model card. Terminal-Bench 4.0 was run in the Claude Code harness, DeepSWE in Mini-SWE-V2, ZeroBench as Pass@5, and the HLE figures in parentheses are the text-only subset.

Three honest observations:

• Every "how much better" number except AA's is DeepSeek's. The 31.2 to 42.3 jump on Terminal-Bench 4.0 is DeepSeek's harness and DeepSeek's run. The one independent cross-check we have is for the September build: AA's own Terminal-Bench 4.0 run scored it 27%, the figure we printed in our Muse Spark Open post, against DeepSeek's 31.2 for the same build. We could not read AA's 1017 figure on that test today, so it stays a dash. AA's six index points are the independent check on the whole table.
• The competitor columns are a generation behind. DeepSeek's card also prints Kimi K3, GLM-5.3, GPT-5.6 Sol, and Claude Opus 5 beside its own rows. Those are DeepSeek's runs or DeepSeek's citations, and two of the four comparators have since been replaced by their vendors. We leave them out.
• "Reduced the model's task runtime and output length" is DeepSeek's claim, and AA's data supports it. Sixteen percent fewer output tokens at about 2% lower per-token speed is the measured version of that sentence.

The Price Everyone Prints Is the Peak Price

AA lists DeepSeek V4.1 Flash 1017 at $0.30 in and $1.20 out per 1M tokens with a 98% cache discount, "based on DeepSeek's API." Those numbers match DeepSeek's peak row exactly. DeepSeek's pricing page also prints an off-peak row, $0.15 cache miss, $0.003 cache hit, $0.60 output, under the rule "Off-peak rates are half of the peak rates."

Peak is narrow: 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday through Friday, excluding Chinese public holidays. Weekends and Chinese holidays are off-peak in full. Converted to Central time, which is our arithmetic, peak today (CDT) runs 8 to 11 PM Sunday through Thursday and 1 to 5 AM Monday through Friday. After US clocks fall back on November 1 (CST) it becomes 7 to 10 PM and midnight to 4 AM. Every US business hour is off-peak.

Two consequences. First, a US team running DeepSeek Flash during the day pays half of what every leaderboard that copies AA's figures shows. If every token were billed off-peak, AA's $0.26 per task would be roughly $0.13, by our arithmetic, since every price component halves; AA does not say which tier its cost figure uses, so treat the pair as a ceiling and a floor, not a quote. Second, the overnight batch job is the one that lands in the expensive window. A Sunday-to-Thursday evening in the US is a weekday morning in Beijing.

The Search Tool Starts Billing at 11 AM CT on October 31

The same October 17 changelog opened DeepSeek's in-house search engine as standalone Search and Fetch APIs, each at $0.75 per 1,000 successful requests. The sentence that matters is on the pricing page: "When you use the DeepSeek API through DeepSeek Harness, Claude Code, or Codex, server-side tool calls already hit the Search and Fetch APIs. These calls are currently free of charge; per-request billing starts at 00:00 Beijing time on November 1, 2026."

Midnight in Beijing on November 1 is 16:00 UTC on October 31, which is 11:00 AM Central on Saturday, October 31. If you point Claude Code or Codex at DeepSeek's endpoint and let the agent search, the calls that are free this week start costing $0.75 per thousand at that hour, with no change to your code. DeepSeek's description of the engine, "hundreds of millions of search requests" a day over "hundreds of billions of web pages," filtering "SEO spam, GEO manipulation, and low-quality AIGC," is DeepSeek's description.

552B and 763B Are Both Right

Two parameter counts for this model are in circulation, and both are DeepSeek's. The V4.1 tech report says the model has "552B backbone parameters and 196B Engram parameters, activating 8B parameters per token during prefill and 16B during decode." The Hugging Face safetensors metadata for the 1017 repo totals 763,205,315,794 parameters.

The config file explains the gap. Beyond the backbone, the download carries two Engram lookup tables and a small draft block used for speculative decoding. By our arithmetic from that file, the Engram tables come to about 196.6B parameters, matching the report's 196B, and the draft block to about 13.6B; subtract both from 763.2B and about 553B remains, the backbone. So 552B is the backbone and is what AA lists; 763B is everything in the download. "16B active" is the decode figure; prefill activates 8B. Self-hosters plan for the 763B one.

Where You Can Run It

DeepSeek's API as deepseek-flash, with vision and thinking mode on by default and a 2,500-session concurrency limit. The web and mobile apps, already switched. The weights on Hugging Face under MIT, which the license says covers "this repository and the model weights." AA benchmarked four third-party hosts on October 18, Baseten, Fireworks, Parasail, and DeepInfra, and lists five API providers in all; we did not check their prices or whether any of them still serves the September checkpoint, so ask before you assume. For the rest of the Chinese open-weight line, start with our Kimi K3 post.

Verified vs Unconfirmed: The Scorecard

Who Should Use It

Use DeepSeek V4.1 Flash 1017 if you want the second-strongest large open-weight model on AA's index at a fraction of a cent per thousand tokens, you run during US business hours or on weekends (all off-peak), and you can re-run your evals this week to confirm the new build behaves the way your prompts assume. For coding and automation agents, DeepSeek's own table says the gains are largest there, and AA's cost-per-task figure says you pay less to find out.

Skip it if you need a model that does not change under a fixed name: the alias moved twice in five weeks, and the only pin is self-hosting a 763B download. Skip it for overnight US batch work priced on the off-peak rate, because the US evening is Beijing's peak. Skip server-side search through Claude Code or Codex after October 31 unless you have counted the calls. And keep an output-token budget; 210M tokens to run AA's index is still half again the class median.

The frontier is elsewhere: the closed flagships at the top of our Index play in a different league at a different price, and even inside the large open-weight class one model still sits above this one on AA's count. What changed this week is the floor: DeepSeek V4.1 Flash 1017 at 45 is now six points above Meta's Muse Spark Open at 39 and seven above Mistral Large 4 at 38, both of which were compared against the September build. Our model comparison has the current standings side by side.

Update, October 25, 2026

• DeepInfer. DeepSeek has open-sourced DeepInfer under Apache 2.0, the offline inference engine behind Flash. It is limited to NVIDIA SM100 or Huawei Ascend 950, and its README says a complete service still needs a KV cache system and data-parallel load balancing that are not open. The self-hosting pin described above now has a reference engine, with that hardware list and those two gaps.
• The rank line is a tie, and it has a date. The "second of 121" in earlier versions of this post is a tie for second with GLM-5.3 at 45; the text above now says so. Moonshot's Kimi K3.1 Pro already scores 55 on the same index as a closed API, with weights promised by November 4. When those weights ship, every rank line in this post moves.

Sources

• DeepSeek API changelog
• DeepSeek models and pricing
• DeepSeek API quick start
• Hugging Face: DeepSeek-V4.1-Flash-1017
• Hugging Face: DeepSeek-V4.1-Flash-1017 config.json
• Hugging Face: DeepSeek-V4.1-Flash, the September build
• DeepSeek V4.1 technical report (PDF)
• Artificial Analysis: DeepSeek V4.1 Flash 1017
• Artificial Analysis: DeepSeek V4.1 Flash (September)
• Artificial Analysis changelog
• GitHub: deepseek-ai/DeepInfer

*Not sure whether an open model that changes under its own name belongs in your stack? Take the free 2-minute quiz and get matched. Then read about the US open model this one now outscores by six points: Muse Spark Open.*