On October 17, 2026, DeepSeek shipped V4.1-Flash-1017 on its website, app, and API under the name deepseek-flash, and published MIT weights. The card lists 763 billion parameters and a 1 million token context.
At a glance
Size and license are from the October 17 model card. The API name, output limit, and prices are from DeepSeek's pricing page. The active split is from the September 10 post about the previous checkpoint.
Total parameters
763 billion
Active parameters
Not restated on this card
Prior checkpoint
552 billion, 8 billion active on input, 16 billion on output
Context
1 million tokens
Max output
384K tokens
License
MIT
Modalities
Image and text in. Vision is listed on the pricing page.
API name
deepseek-flash
Peak price
$0.30 / million input tokens on a cache miss, $1.20 / million output tokens
Off-peak
Half of peak
How to read this claim
Measurement — This card lists 763 billion parameters and does not say how many are active.
Bill — Peak price: $0.30 / million input tokens on a cache miss, $1.20 / million output tokens
Door — No independent rerun of the benchmark table is cited.
All three columns are from the October 17 model card. Opus 5 was not rerun for this page.
Prior Flash
1017
Opus 5
Terminal-Bench 4.0
31.2
42.3
51.8
DeepSWE v1.1
74.2
77.0
74.0
HLE, no tools
36.8
42.6
56.3
ProgramBench
20.3
29.1
37.0
What changed
The previous Flash checkpoint was the file behind deepseek-flash. DeepSeek says this one is live on the website, the app, and the API under that same name, finishes tasks faster, and writes shorter outputs. The pricing page lists peak cache-miss input at $0.30 per million tokens and output at $1.20. Off-peak is half. The card's Terminal-Bench 4.0 row moves from 31.2 to 42.3.
What it introduces
A newer MIT checkpoint on the existing deepseek-flash endpoint, plus weights. DeepSeek also opened Search and Fetch at $0.75 per 1,000 calls.
Who benefits Anyone who already calls deepseek-flash, and anyone choosing between that endpoint and an open checkpoint they would serve themselves.
Try this
If you already call deepseek-flash, run one job you already score and compare the answer and the bill. Peak cache-miss input is $0.30 per million tokens and output is $1.20, on DeepSeek's pricing page. Off-peak is half. Download the weights only if you need to serve the file yourself. The benchmark deltas are DeepSeek's.
What we do not know
This card lists 763 billion parameters and does not say how many are active.
The September 10 post for the previous Flash checkpoint described 552 billion parameters, with 8 billion active on input and 16 billion on output. This card does not say that split still applies.
No independent rerun of the benchmark table is cited.
DeepSeek says tasks finish faster and outputs are shorter. The October 17 post does not say the token price changed.
Same day
DeepSeek also opened Search and Fetch. The pricing page lists $0.75 per 1,000 successful calls for each. Calls through DeepSeek Harness, Claude Code, or Codex stay free until 00:00 Beijing time on November 1, 2026, which is 16:00 UTC on October 31.
Evidence
PrimaryAnnouncement. The October 17 announcement and the model card are DeepSeek's. The benchmark table is DeepSeek's. Prices are from DeepSeek's pricing page.
Independent No independent rerun is cited. The benchmark table, including the Opus 5 column, is DeepSeek's.
On October 17, 2026, DeepSeek published V4.1-Flash-1017 and said it is live on the website, the app, and the API. The pricing page names the model deepseek-flash and lists the version as DeepSeek-V4.1-Flash-1017, with a 1 million token context and a 384K maximum output. Peak cache-miss input is $0.30 per million tokens and output is $1.20. Off-peak is half. The model card lists 763 billion parameters and the MIT license. In DeepSeek's table, Terminal-Bench 4.0 moves from 31.2 to 42.3, and Opus 5 is listed at 51.8. DeepSWE v1.1 moves from 74.2 to 77.0, with Opus 5 at 74.0. DeepSeek says the Terminal-Bench run used the Claude Code harness and DeepSWE used the Mini-SWE-V2 harness.
Why it matters
Most readers will meet this as an API name, not as a 763 billion parameter download. deepseek-flash now points at this checkpoint, at the price on DeepSeek's pricing page. On DeepSeek's table, Terminal-Bench 4.0 rises from 31.2 to 42.3 and Opus 5 is still at 51.8, while DeepSWE v1.1 has this checkpoint at 77.0 and Opus 5 at 74.0. Those are DeepSeek's measurements. Beam, released the day before, is Apache 2.0 with 23 billion active parameters. Muse Spark Open is under Meta's license with 42 billion active. This card does not restate an active count.
DeepSeek's table also lists HLE without tools at 42.6 for this checkpoint and 36.8 for the previous Flash, with Opus 5 at 56.3. The card marks a text-only subset at 44.0 and 39.1. Those figures are DeepSeek's.
The card says this release does not include a Jinja chat template. It points to a Python reference encoder and a separate Rust toolkit for the prompt format.
The September 10 post said the previous Flash checkpoint's KV cache needed one quarter of the HBM and one eighth of the SSD storage of the generation before it. That sentence describes the previous checkpoint.