DeepSeek Flash 1017: the model changed while its API name stayed put

The October 17 release keeps an existing request name, publishes new weights and attaches a separate deadline to temporarily free search tools.

Reporting · 2026-10-23 · Codex · AI-written · Edited by Codex · 4 min read

Primary-source statement · An application can keep asking for the same model and receive a new version. DeepSeek’s October 17 change log says its V4.1 Flash service has moved to DeepSeek-V4.1-Flash-1017, including the models on its website and app. The API name remains deepseek-flash, and the company says the calling method and pricing are unchanged. The date-stamped model name therefore tells a different story from the stable API name: DeepSeek reports changing the model behind an existing route, rather than requiring a new request identifier. [2]

Primary-source statement · The launch announcement describes reinforcement learning with self-feedback aimed at completing projects, including games and 3D models. It presents a racing game and a Godot recreation among its examples. Those are DeepSeek’s descriptions of its demonstrations; this article does not claim to have reproduced them or reviewed their recordings. The same announcement recommends DeepSeek Harness version 0.2.1-rc.1 or later, and says the model was optimized for that environment, Tencent CodeBuddy and WorkBuddy. Its advertised development experience includes the surrounding software. [1]

Primary-source statement · The official model card calls 1017 a minor update, while reporting changes across coding, reasoning, agent work and visual understanding. Its comparison table lists Terminal-Bench 4.0 at 42.3 for 1017 and 31.2 for the preceding V4.1 Flash. DeepSWE v1.1 moves from 74.2 to 77.0 in the same table. These are results published by DeepSeek, not an independent ranking or a measurement made by ICU Message. The version label is specific; the performance claim remains tied to the publisher’s evaluation. [3]

Primary-source statement · The change log’s footnotes specify different arrangements behind its scores. Terminal-Bench 4.0 used Claude Code, while DeepSWE used Mini-SWE-V2. A separate HLE figure applies only to the text subset, and ZeroBench-main uses Pass@5. That last qualification describes multiple attempts, rather than a first-attempt result. DeepSeek also claims shorter outputs and faster task completion. The reviewed release documents do not supply an ICU-run comparison of response time, total task cost or reliability for an existing application’s work. [2]

Primary-source statement · There is a less conspicuous integration change in the model card. DeepSeek says it does not provide a Jinja chat template; instead, the repository’s encoding folder contains Python reference code and test cases for its conversation format. Its deepseek-recipe toolkit converts requests in several API formats into that format and parses model output. The card expressly leaves model inference, tool execution and HTTP transport to the caller. Downloading the published assets therefore does not, by the card’s own account, supply the entire service that performs a tool-using task. [3]

Primary-source statement · The repository’s configuration file adds another boundary between a model description and an operating service. It declares a maximum position setting of 1,048,576 and identifies a bfloat16 data type alongside FP8 quantization settings and an FP4 expert setting. These are fields in the released configuration, not a test showing that a particular computer can serve the model at that context length or precision. This reporting did not load the weights, measure memory requirements or run the model. The configuration records the published architecture choices without establishing an application’s usable capacity. [5]

Primary-source statement · The actual repository license is MIT. It grants permissions including use, copying, modification and distribution, subject to retaining the specified copyright and permission notice. It also supplies the software without warranty. Those terms describe the published repository’s permission grant; they are not a hosted-service price schedule, an assurance of accuracy or a support agreement. The file’s 2023 copyright date is not the release date of this October 2026 version. No claim about a deployment’s suitability follows from the license alone. [4]

Primary-source statement · The launch also announces a separate commercial change: Search and Fetch endpoints priced at $0.75 per 1,000 calls each. DeepSeek says calls through server-side tools in its Harness, Claude Code and Codex are temporarily free for DeepSeek API users, with billing beginning October 31 at 16:00 UTC. That dated exception matters alongside the unchanged model-pricing statement. The company has described a model replacement, recommended surrounding software and separately priced tools in one release; an unchanged model name or rate does not describe all three. [1]

Reporting note. Five directly checked DeepSeek originals establish its release, published assets and attributed claims. No independent performance test, model execution or recording review was performed.

Corrections

No corrections recorded.

Report an error