DeepSeek 1017: the old model name now serves new weights, and Artificial Analysis measures almost half the cost per task

On 17 October 2026 DeepSeek replaced the weights behind its Flash model. The model name in the interface stayed, the call stayed, and the documentation names no price change. The DeepSeek documentation says it plainly: deepseek-flash remains the model name to use, that name is now served by DeepSeek-V4.1-Flash-1017, and the calling method is unchanged. The earlier names deepseek-v4-flash and deepseek-v4-flash-vision-exp are still accepted, the models behind them are retired, and requests are billed at the Flash price (API documentation start page, announcement of 17 October: DeepSeek-V4.1-Flash-1017 Release).

The weights have been on Hugging Face since the same day at 02:40 UTC: 83 files, 510.3 GB, 48 safetensors, MIT licence (model card). What started our recheck was a YouTube video by Donato Capitella from 18 October 2026, in which he runs the checkpoint on AMD unified-memory machines (DeepSeek V4.1 Flash 1017 Update: 44% Fewer Tokens, Running on Strix / Gorgon Halo with DwarfStar). Figures that appear only there are marked "per the video".

A silent model swap happens when a provider replaces the weights behind an unchanged model name in its API. The call stays the same, while answers, token consumption and runtime can change. Anyone who recorded measurements against that name is afterwards comparing against a different model without any change in their own configuration.

Updated on 5 November 2026: we reopened every source, replaced a wrong documentation link, corrected the benchmark details and the pricing windows, fixed the file count of the repository to 83, and added a postscript with our own measurements. The Artificial Analysis figures date from 19 October and can no longer be checked today (see the section on them).

What the cards say

The architecture is documented on the card of the predecessor V4.1-Flash (V4.1-Flash model card): a backbone of 552 billion parameters, 8 billion active for input and 16 billion for output, plus Engram tables of 196 billion parameters that are only touched through token lookup. Each MoE layer has 1 shared and 384 routed experts, 6 of them active per token. The global KV cache is packed in FP4 at 890 bytes per token, roughly a quarter of DeepSeek-V4-Flash according to the card. Context is 1 million tokens (pricing page).

The card of 1017 lists benchmarks, all of them vendor figures: Terminal-Bench 4.0 at 42.3 against 31.2 for V4.1-Flash, Humanity's Last Exam without tools at 42.6 against 36.8, ProgramBench at 29.1 against 20.3, DeepSWE v1.1 at 77.0 against 74.2. The card names Claude Code as the harness for Terminal-Bench and Mini-SWE-V2 for DeepSWE. Temperature and number of runs are not stated. DeepSeek phrases the result as a better position on the cost-performance frontier across several agent harnesses.

Measured independently: same speed, fewer tokens

On 19 October 2026 Artificial Analysis carried both checkpoints in the same index (version 4.3.2, Max variant): intelligence index 45 against 39, cost per task 0.26 against 0.49 US dollars, output tokens across the whole index 210 million against 250 million, output speed 227 against 231 tokens per second. That works out to 16 percent fewer output tokens and 47 percent lower cost per task (1017, V4.1-Flash).

We could not re-check these values on 5 November. The site now scores with index version 5.0.1 and shows 1017 at an intelligence index of 33, 174 million output tokens, 175 tokens per second and 0.60 US dollars per task, and no value at all for V4.1-Flash. The two states are not comparable. We could not retrieve an archived copy of the 19 October page. Treat the October figures as a statement by Artificial Analysis as of 19 October.

We did not measure the two models against each other for this table. Our own measurements are in the postscript below.

Two things stand out in the October figures. Speed is essentially unchanged, the gain sits in the output. And cost fell 47 percent while output volume fell only 16 percent. The pricing page lists 0.30 US dollars per million input tokens without cache hit and 1.20 US dollars per million output tokens at peak today. It does not say whether the token price changed with the swap, and in the figures we had, Artificial Analysis does not explain the gap between 16 and 47 percent. Our guess, not evidenced: a larger part of the drop could come from input and cache segments, meaning fewer rounds in which the same context is sent again. For agents the number of rounds sets the price. How to work that out for your own load is in What an AI model really costs, and related assessments are collected in the topic hub AI explained.

What an unchanged model name means in operations

For a test baseline this is a silent cut. Everything measured in September against deepseek-flash is now compared against a different model without a single configuration file having been touched. DeepSeek documents the change openly but names no way to postpone it.

If a fixed version is needed, it has to be obtained elsewhere. OpenRouter lists deepseek/deepseek-v4.1-flash-1017 as its own model next to the older deepseek/deepseek-v4.1-flash (model page). Or host it yourself. Our recommendation for every lane that points at a foreign endpoint: write the model reference from the response into the log line, otherwise a regression noticed three weeks later can no longer be attributed to a swap. On the pricing page, the half rate applies outside the peak windows, and those are 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays (excluding Chinese public holidays). All other hours, weekends included, are off-peak (pricing). If you want help running agents on this basis, our services show what we offer.

Locally this is a memory question

This post covers the silent swap of weights behind the name deepseek-flash and the figures around it. Why V4.1-Flash did not fit on two DGX Spark in October is in Two DGX Spark in our own test, and how large the Pro model is on disk in DeepSeek V4 Pro measured.

The channel runs the model with the DwarfStar engine by Salvatore Sanfilippo (antirez), the project antirez/ds4 with about 24,200 stars on 5 November 2026 (repository). The finished Q2 GGUF file weighs 365.7 GB. The companion file in the quantization repository of kyuz0 gives the split: the tested runtime keeps 151.76 GiB of language-model tensors in memory before context and buffers, and the file adds 188.83 GiB of Engram data on a fast local SSD. Tested on AMD gfx1151 under ROCm (quant repository). The matching 1017 draft model for speculative decoding (DSpark, 7.42 GiB) cannot be combined with SSD streaming of the experts according to the same file. antirez uploaded his own 1017 file on 19 October at 07:57 UTC, followed by nine MXFP4 parts between 08:34 and 09:15 UTC (repository).

The technique behind the speed is called SWA Bounded Replay at DeepSeek and is described on the V4.1-Flash card: missing SWA KV states are rebuilt by replaying only the most recent tokens of the window instead of keeping them on SSD. The persistent KV cache drops to about one eighth of DeepSeek-V4-Flash according to the card. As an implementation for DwarfStar it sits in an open pull request in the fork kyuz0/ds4, which the video description links (pull request 2). It is not the antirez main branch and not merged. The numbers in it come from the author of the pull request (GitHub account kyuz0), on Strix Halo with V4.1-Flash Q2. We infer that this account belongs to the channel from the fact that the video description links the pull request; the pages do not say so explicitly:

The pull request names its own limits: off by default, resident execution only, qualified on ROCm only, prefill logits change by design.

Our arithmetic at the start of October was: the server with four RTX 5090 has 4 × 32,607 MiB, about 127 GiB of graphics memory in total, and 91 GiB of system memory. 151.76 GiB of tensors does not fit into graphics memory alone. Machines with a large unified memory are the natural target, as AMD's Ryzen AI Max line provides it (product page of the Ryzen AI Max+ PRO 495). We did not check whether a split across graphics and system memory runs with DwarfStar on Blackwell using this file. The postscript below shows that it works with a different quantization.

The number from the video and where it stands

According to the video description, Capitella reports 44 percent fewer generated tokens and 42 percent less agent time over eight matched tasks, as an early result of a Terminal Bench Mini run that was still going. A repository of the same name with the Core-19 suite sits under the GitHub account kyuz0 (Terminal Bench Mini); that it is the harness from the video we infer from the name. Its README does not mention 1017. Whether the committed result files contain 1017 runs we could not check, because they are only reachable through a web view. The 44 percent is therefore an observation from a run in progress, not a published measurement. The independent 16 percent of Artificial Analysis stands next to it.

For context on the channel: the preceding video of the series carries "#AMD_Partner. This video is sponsored by AMD" in its description (video). The 1017 video does not carry that note in its description, but the channel tests AMD hardware.

Postscript of 5 November: a DeepSeek 1017 runs here

Since 4 and 5 November 2026 we run DeepSeek-V4.1-Flash-1017 ourselves. We use the quantization ISTA-DASLab/DeepSeek-V4.1-Flash-1017-GSQ-RCO-GGUF at 2.25 bits per weight (set IQ2_XS), and the runtime is Strata. We do not use the Q2 file from the video. The statement "not with what is in the building" from the first version of this post no longer holds.

These are single runs and values may differ on repetition. We did not measure quality on German mail and ERP workloads, long-term operation or behaviour under concurrent load. We did not verify the vendor claims about the quantization quality (ISTA-DASLab, Strata).

What we take from it

The move to our own weights changes the question. What is open now is whether the model is good enough on German workloads and whether 85 tokens per second suffice for agents. For the choice between self-hosting and the API we recommend counting generated tokens per completed task, together with the number of rounds, for four weeks. If a model change moves that figure by something like the Artificial Analysis numbers, the decision it triggers is a procurement one that no longer needs estimating. How we run smaller models in house is in Qwen3.8-27B Self-Hosted, and we worked out the price per token on two cards in Radeon R9700 against RTX 5090.

Further reading

• DeepSeek-V4.1-Flash-1017 Release, announcement of 17 Oct 2026
• DeepSeek API documentation, start page with the note on the retired names
• DeepSeek API pricing page with peak and off-peak windows
• Model card DeepSeek-V4.1-Flash-1017 on Hugging Face
• Model card DeepSeek-V4.1-Flash with architecture, Engram and KV cache sizes
• Artificial Analysis: DeepSeek V4.1 Flash 1017
• Artificial Analysis: DeepSeek V4.1 Flash
• Video by Donato Capitella, 18 Oct 2026: DeepSeek V4.1 Flash 1017 Update
• Preceding video, sponsored by AMD: DeepSeek V4.1 Flash on Strix Halo
• Pull request 2 in the fork kyuz0/ds4: decoder SWA bounded replay
• DwarfStar (antirez/ds4), README
• Quantization repository kyuz0/deepseek-v4.1-flash-gguf
• Quantization repository antirez/deepseek-v4.1-flash-gguf
• ISTA-DASLab: DeepSeek-V4.1-Flash-1017 GGUF
• Terminal Bench Mini, harness and Core-19 suite
• OpenRouter: deepseek-v4.1-flash-1017
• AMD Ryzen AI Max+ PRO 495 product data