DeepSeek 1017: the old model name now serves new weights, and the cost per task almost halves

On 17 October 2026 DeepSeek replaced the weights behind its Flash model. The model name in the interface stayed, the price stayed, the call stayed. The documentation for api.deepseek.com 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 the requests are billed at the Flash rate (API documentation). No date appears in the model name there.

The weights have been on Hugging Face since the same day at 02:40 UTC: 88 files, 510.3 GB, 48 safetensors, MIT licence, 17,924 downloads (model card). What started our recheck was a video by Donato Capitella from 18 October, in which he runs the checkpoint on AMD unified-memory machines and reports a number we could not find published anywhere.

What the cards actually say

The backbone is 552 billion parameters, 8 billion active during prefill and 16 billion during decode, plus an Engram table of 196 billion parameters that is only touched sparsely (predecessor card). Context is 1 million tokens. One shared and 384 routed experts are available per token, six of them active. The global KV cache is packed in FP4 at 890 bytes per token, roughly a quarter of the predecessor. The half trillion from the video is a fair rounding.

The card also lists benchmarks, all of them vendor figures: Terminal-Bench 4.0 at 42.3 against 31.2 for the previous checkpoint, 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. Harness, temperature and sample count are not stated. The improvement is phrased as a better position on the cost-performance frontier across several agent harnesses.

Measured independently: same speed, fewer tokens

Artificial Analysis carries both checkpoints in the same index, read on 19 October: 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 (1017, predecessor).

US dollars per task of the Artificial Analysis Intelligence Index v4.3.2, Max variant, read on 19 October 2026. Token prices on both sides sit at 0.30 for input and 1.20 for output per million. (Quelle: Artificial Analysis)

Two things stand out. Speed is essentially unchanged, the gain sits elsewhere: in the output. And cost fell 47 percent while output volume fell only 16 percent. The token prices are identical, so most of the saving has to come from the input and cache segments, which means fewer rounds in which the same context is sent again. That is an inference; Artificial Analysis does report token types per segment. For operations it is the more useful half of the story, because agent workloads are priced by rounds rather than by speed (What one task costs).

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, it is just not a change anyone can decline or postpone.

If a version is needed, it has to be obtained elsewhere. OpenRouter lists deepseek/deepseek-v4.1-flash-1017 as its own model and keeps the older one as well. Or host it yourself. Concretely for us: in every lane that points at a foreign endpoint, the model reference from the response belongs in the log line, otherwise a regression noticed three weeks later can no longer be attributed to a swap. A side detail on the same pricing page: nights and weekends are billed at half rate, the peak windows are 01:00 to 04:00 and 06:00 to 10:00 UTC on working days (pricing).

Locally this is a memory question

The channel runs the model with the DwarfStar engine, the project of Salvatore Sanfilippo with 23,937 stars (repository). The finished Q2 GGUF file weighs 365.7 GB. The companion file in the quantization repository gives the split: 151.76 GiB have to stay resident as tensor spans, 188.83 GiB of Engram tables run from SSD, target hardware gfx1151 under ROCm (quant repository). On top comes an 8 GB draft model for speculative decoding, which according to the same file cannot be combined with SSD streaming.

The technique behind the speed is called SWA Bounded Replay at DeepSeek and is documented on the model card as the vendor's own inference method: only the tokens in the last window pass through the decoder layers, which drops the persistent KV share to about one eighth. As a port into the engine it sits as an open, unmerged pull request with tables in the same range as the video: resident 215.84 to 422 tokens per second exact against 393 to 443 with bounded replay, decode around 14.6, with the draft model 17.66 to 18.34 (pull request #2). The limits are named in the request itself: off by default, qualified for ROCm only, prefill logits changed by design, weighted 0.197360214 against 0.195660362.

Our own arithmetic is short and measured: the largest inference machine in the building has four cards at 32,607 MiB each, 91 GiB of system memory and 1.7 TB of free disk. 151.76 GiB of resident spans does not fit there, even after splitting across the cards. This class of model is addressed through 128 to 192 GB of unified memory, which is what AMD builds the Ryzen AI Max+ PRO 495 with its 256-bit bus for. Our real size remains the 27B model in self-operation (recomputed in October).

The number from the video and where it stands

Capitella reports 44 percent fewer generated tokens and 42 percent less time per task, measured over the first 8 tasks of a run through Terminal Bench Mini that was still going. The harness and its committed results live in one repository: suite Core-19, 19 tasks, 21 runs, 4 platforms, and the string 1017 does not occur once in the results index read on 19 October (suite and runs). The 44 percent is an observation from a run in progress, published nowhere, and the independent 16 percent stands next to it.

Three further points, as of 19 October. The video says antirez had not published quants for 1017 yet and that he quantized them himself; at 07:57 UTC on 19 October antirez/deepseek-v4.1-flash-gguf received a 1017 file, and nine MXFP4 shards followed between 08:34 and 09:15 UTC. The channel is AMD-partnered, the preceding video in the series carried the tag #AMD_Partner. And an aggregator reports regressions in knowledge and long context without any measurement of its own; that stays open, while the independent figures point the other way.

What we make of it

We looked at our own lane first: the gateway configuration on our LiteLLM host carries no DeepSeek provider at all, all six model names there point at our own machines, and nothing in our vLLM and llama.cpp production changes because of this checkpoint (llama.cpp in house operation).

The second conclusion is a refusal to buy. Our four cards do not run this model, a machine with 128 or 192 GB of unified memory would, and 0.26 US dollars per task at DeepSeek is cheaper than any operating calculation we could build before a machine is even standing. Anyone who wants this size class in house buys memory before speed; we did that arithmetic on 16 October already (R9700 against RTX 5090).

What we will do instead costs nothing: count generated tokens per completed task across our own lane for four weeks, together with the number of rounds. If a model change moves that figure by anything like the independent measurement, the decision it triggers is a procurement one, and we will have the baseline for it in the log.