DeepSeek Ships V4.1-Flash-1017: Faster, Leaner 763B Model

DeepSeek has released DeepSeek-V4.1-Flash-1017, an update to its workhorse Flash model that the company says completes tasks faster and produces shorter outputs than the version it replaces, while scoring higher on coding and agentic benchmarks. The model went live on October 17 through DeepSeek's website, app, and API, with open weights published on Hugging Face under an MIT license. It is a 763-billion-parameter multimodal model with a context window of up to one million tokens, and it is now the model served whenever a developer calls the deepseek-flash endpoint.

What's new in DeepSeek-V4.1-Flash-1017

DeepSeek frames the release as "a Pareto improvement in performance and cost" over the version it replaces, meaning it claims to be better on benchmark scores, speed, and cost at the same time rather than trading one for another. According to the company's October 17 release notes, the gains come from additional reinforcement learning with self-feedback, a training approach aimed at improving how the model handles long, multi-step "project development" tasks — the kind of work where a model has to plan, write, run, and debug code across many turns rather than answer a single prompt.

DeepSeek illustrated the update with demo projects rather than a spec sheet: a 3D model of Beijing's Forbidden City, a browser-playable racing game, a recreation of a Portal 2 level built in the Godot engine, and a historical 3D scene called "Valley of Ages." The release notes also point to a chart plotting DeepSWE v1.1 Pass@1 score against cost per completed task, measured with the mini-swe-v2 evaluation harness, as evidence that the new build gets more done per dollar. DeepSeek says the model is tuned to work with its own DeepSeek Harness agent framework (it recommends version 0.2.1-rc.1 or later), as well as with Tencent's CodeBuddy and WorkBuddy.

The model card for the update, hosted on Hugging Face, describes it more plainly as "a minor update" that "improves reasoning, coding, agentic tasks, and visual benchmarks, while reducing task execution time and output length." That last point — shorter outputs — matters for cost control on a model billed by the token, since a model that reaches the same answer in fewer tokens is cheaper to run even before any price change.

Size, context window, and modality

DeepSeek-V4.1-Flash-1017 carries 763 billion parameters in total, matching the count Hugging Face lists for the original V4.1-Flash it replaces — this is a retrain and refinement of the same architecture, not a bigger model. Of that total, the model's technical paper attributes 552 billion parameters to its core Mixture-of-Experts backbone, with the remainder going to a 196-billion-parameter "Engram" conditional memory module. Only a fraction of the backbone is active for any given token: 8 billion parameters fire during prefill (reading the prompt) and 16 billion during decode (writing the response), which is how a model with hundreds of billions of stored parameters can still run at Flash-tier speed and price.

The model is tagged image-text-to-text on Hugging Face, meaning it accepts interleaved images and text as input and responds in text — native visual understanding rather than a bolted-on vision add-on. Its context window extends to 1 million tokens, and the model card recommends configuring at least 256,000 max output tokens for agentic or long-document workloads, while DeepSeek's pricing documentation caps actual output length at 384,000 tokens per response.

Inside the KV cache squeeze

The headline engineering work behind this model family isn't the October update itself but the architecture it's built on, detailed in DeepSeek's paper "DeepSeek-V4.1-Flash: Pushing the Limits of KV Cache Compression." Long-context, multi-turn agent work is bottlenecked less by raw compute than by the key-value (KV) cache — the running memory of everything a model has already processed in a conversation, which has to stay resident in fast GPU memory (HBM) or spill to slower storage as conversations stretch on.

DeepSeek's answer is a Causal Encoder-Decoder design with 40 total layers split evenly between encoder and decoder stacks, paired with a technique the paper calls Compressed Sparse Attention 2 (CSA2) that lets layers reuse cached keys and values instead of recomputing them, plus storage of the cache itself in 4-bit floating point (FP4) rather than higher precision. The combined effect, per the paper, is a global KV cache footprint of about 890 bytes per token in HBM — roughly a quarter of what the prior-generation DeepSeek-V4-Flash required. A separate optimization called SWA Bounded Replay cuts the footprint that has to persist to SSD or host memory to roughly an eighth of that older model's requirement. DeepSeek says quality held up "substantially better" than the baseline despite the smaller cache, though the paper's public abstract doesn't publish the specific accuracy figures behind that claim. The model was pretrained on 45 trillion multimodal tokens, with context length extended out to the full 1 million tokens during a later 34-trillion-token stage of training.

Benchmark scores, and what's actually comparable

DeepSeek's model card for the 1017 build lists results across a wide set of coding and agentic benchmarks, including a DeepSWE v1.1 score of 77.0 and a Humanity's Last Exam (HLE) text-only score of 44.0. Both of those benchmarks also appear, under the same names and versions, on the original V4.1-Flash's model card from September — making them the only two scores in this release that can be compared like-for-like. On DeepSWE v1.1, the score rose from 74.2 to 77.0. On HLE (text-only), it rose more sharply, from 36.8 to 44.0.

Several other suites DeepSeek highlights for the 1017 build — including Terminal-Bench 4.0 and Automation-Bench v1.0.6 — replace differently versioned benchmarks the September model was scored on (Terminal-Bench 2.1 and an earlier AutomationBench), so those pairs aren't a clean apples-to-apples comparison even though the newer scores look strong in isolation. The table below lists the 1017 model's published scores as reported on its own model card.

Pricing and API access

DeepSeek-V4.1-Flash-1017 is served under the existing deepseek-flash model name, and DeepSeek's API pricing page shows no price change attached to this specific update — the October notes describe the release purely as a quality and speed improvement at the existing price. Two legacy model names, deepseek-v4-flash and deepseek-v4-flash-vision-exp, are also still accepted and are billed at the same Flash rate, routing silently to the new build. DeepSeek runs time-of-day pricing, with peak hours (01:00–04:00 and 06:00–10:00 UTC, Monday through Friday, excluding Chinese public holidays) priced at double the off-peak rate.

Prices are per million tokens. DeepSeek also charges for its web Search and Fetch tool calls at $0.75 per 1,000 successful requests per endpoint; calls made through DeepSeek Harness, Claude Code, or Codex remain free of charge until per-request billing begins on October 31 at 16:00 UTC.

What happened to V4-Pro

This update continues a routing change DeepSeek made alongside the original V4.1-Flash launch in September: since September 14, any API call naming deepseek-v4-pro has been silently redirected to V4.1-Flash and billed at Flash's lower rate, rather than hitting the larger 1.7-trillion-parameter V4-Pro model. DeepSeek said at the time of the original V4.1-Flash launch that the routing would stay in place "until V4.1-Pro launches," without giving a date. As of this release, DeepSeek-V4-Pro-0813 is still listed on the company's Hugging Face organization page, but it is no longer the model actually serving Pro-tier API traffic.

Where to get it

The weights are published on Hugging Face under an MIT license, with a mirror on ModelScope for users in mainland China. DeepSeek's own chat app and website serve the updated model directly, and API access runs through the same deepseek-flash model name developers were already using. The deepseek-ai Hugging Face organization also hosts the inference code, encoding library, and example scripts needed to run the model outside of DeepSeek's own API, along with quantized and fine-tuned community variants.

How it fits alongside rival models

DeepSeek's release lands in a crowded month for frontier model updates. Google has been rolling out its Gemini 4 Argon update, Anthropic shipped a new Opus release that outside writers have already measured against rivals like Fable 5.1, and the image-and-video side of the market has its own fast-moving comparisons, including Grok Imagine against Nano Banana Pro. DeepSeek's pitch is different from most of that pack: rather than chasing the largest possible flagship, it has spent two releases in a row optimizing a mid-sized, open-weight model for the token-cache math that determines how expensive long agent sessions actually are to run — the same bottleneck Chinese and Western labs alike are racing to solve as "agentic" use cases replace single-turn chat as the default way people use these models.

Why this release matters

Model-card point releases rarely make news on their own, and DeepSeek has shipped several quiet "Pareto improvement" updates to its Flash line over the past year without much coverage. What makes the KV cache compression work behind V4.1-Flash notable is that it targets cost, not just capability — a 1-million-token context window is only useful to a developer if the memory needed to maintain it doesn't make long sessions prohibitively expensive, and DeepSeek's own figures put the persistent memory cost of this model line at roughly an eighth of its immediate predecessor. Combined with a 763-billion-parameter model that activates as few as 8 billion parameters per token, the architecture is explicitly built to make long, multi-step, tool-using agent sessions cheaper to run at scale, which is the direction most of the frontier-model market — including the Hugging Face and open-source ecosystem the Hangzhou-based lab publishes into — has been pushing for the past year.

What's next

DeepSeek has not given a release date for V4.1-Pro, the larger model that is meant to eventually replace the routing workaround currently sending V4-Pro-tier requests to Flash. The company's release notes point developers toward DeepSeek Harness version 0.2.1-rc.1 or later to get the most out of the 1017 update's agentic tuning, and toward its Search and Fetch tools ahead of November's end to free-tier access for Claude Code, Codex, and Harness users. Given the cadence of the last two releases — a full architecture change in September followed by a reinforcement-learning refinement five weeks later — a further point update to the Flash line, or the long-awaited Pro-tier model built on the same KV-cache-compressed architecture, looks like the more likely next move from DeepSeek than an entirely new model family.

Written byTheo Park

Theo Park runs the AI desk at Pandromeda. He follows model launches from the frontier labs and the open-weight community, tracks the assistants and developer tools built on them, and explains what each release changes on pricing, capability and safety. His reporting leans on primary sources: model cards, technical reports, API documentation and the companies' own announcements.

More from Theo ParkAll AI coverage