AdvertiseDonate
☰ SectionsNews from America’s biggest cities.
United Media Now · 2000 Auburn Drive, Suite 200, Beachwood, OH 44122

New Deepseek Model Cuts Memory Costs But Struggles With Facts

A recent update to Deepseek's V4.1 Flash model drastically lowers hardware requirements, yet independent tests reveal persistent issues with invented information.

By Abigail Brooks ·

Illustration for: New Deepseek Model Cuts Memory Costs But Struggles With Facts
Illustration generated for this article, not a news photograph.

Deepseek has rolled out a new version of its V4.1 Flash model, designated as build 1017. While the label might suggest a minor patch, the changes appear to significantly improve efficiency. The company markets this release with the phrase wish-driven programming made real, a nod to broad, open-ended coding tasks. However, outside testing suggests the model still has a tendency to fabricate details.

The market for large language models is currently divided into two distinct camps. Commercial closed models from major labs offer high accuracy but come with a steep price tag. In contrast, Chinese open-weight alternatives often sacrifice some reliability in exchange for much lower operating costs. This trade-off is particularly noticeable in long-running agent workloads.

Technically, V4.1 Flash is a Mixture-of-Experts architecture. It contains a total of 552 billion parameters, but only 16 billion are active during any given run. The model is multimodal, handling both text and images. It also supports a context window of one million tokens, which has become a standard expectation in the industry.

In Deepseek's own evaluations, the new model performs between GPT-5.6 Sol and Claude Opus 5. It even outperforms the Anthropic model in specific task tests. The most significant improvement, however, is in efficiency. Compared to the previous V4 Flash, the 1017 variant requires only a quarter of the high-bandwidth memory and an eighth of the solid-state drive storage for its cache. It uses just 890 bytes per token entry.

Deepseek supports these claims with a cost chart based on the DeepSWE September 2 dataset. Both the standard and 1017 versions fall into the efficient top-left zone of the graph. Even low-effort runs of the 1017 model complete tasks at a rate similar to the maximum settings of more expensive competitors. While DeepSWE is just one of many benchmarks, the results are noteworthy.

Pricing remains competitive for the open-weight community. At off-peak hours, Deepseek charges $0.15 per million input tokens and $0.60 per million output tokens. Cache-hit input drops to just $0.003 per million tokens. Since the weights are open, third-party providers can run the model and have historically offered lower prices than Deepseek's own API. The company's search feature costs $0.75 per 1,000 queries, which is cheaper than rivals like Parallel at $1 per 1,000.

Artificial Analysis ranks the 1017 model second among large open-weight options on its intelligence index. It scores 45, whereas the top-scoring Claude Opus 5.5 sits at 58. The model is also among the fastest in its category. Only Xiaomi's MiMo V2.6-Pro beats it in terms of cost per task.

The main weakness is reliability. On the Omniscience benchmark, the model answered 43% of questions correctly. This is good for an open-weight model but lags behind proprietary systems. It hallucinated on 67% of the questions it should have declined to answer. The benchmark penalizes wrong answers and ignores non-answers. The 1017 model scored 5, compared to 20 for Kimi K3-max and at least 42 for frontier models.

This profile makes the model best suited for specific jobs. Tasks with verifiable outputs, such as coding or format conversion with review, work well because agents can check and retry. Open-ended research or precise document drafting is better left to frontier models or humans, where errors are harder to catch. This fast-and-loose behavior seems to be a trend among capable open-weight models. Even Kimi K3, praised for being cheap, produced invented information during short lookups. Despite these flaws, open-weight models remain vital for keeping AI agents affordable and checking price hikes from big labs.

More Tech · Report a business · Got a Story?

Get United Media Now in your inboxNews from America’s biggest cities. Free, a few times a week, unsubscribe any time.

Related stories

Get United Media Now in your inboxNews from America’s biggest cities. Free, a few times a week, unsubscribe any time.