Deepseek's V4.1 Flash 1017 slashes memory needs but struggles with accuracy

Deepseek has launched a new version of its open-weight V4.1 Flash model, designated as 1017. While the release looks like a minor update on the surface, the changes deliver a clear boost in efficiency. The company markets the tool with the slogan wish-driven programming made real. This phrasing suggests the model is built for open-ended tasks that are described in broad, general terms rather than strict instructions. However, independent testing reveals a persistent issue with hallucination that remains a concern for users.

The large language model market currently divides into two distinct groups. Closed commercial models from major labs cost more to operate but tend to produce highly accurate results. In contrast, Chinese open-weight competitors often sacrifice some reliability in exchange for much lower running costs. This trade-off is particularly noticeable in long-running agentic work where efficiency is a priority.

V4.1 Flash is a Mixture-of-Experts model. It contains 552 billion total parameters, but only 16 billion are active during a single run. The system is multimodal and accepts both text and images. It supports a context window of 1 million tokens, a standard that is now common in the industry. Deepseek’s own benchmarks place the model between GPT-5.6 Sol and Claude Opus 5. It even outperforms Anthropic’s model in specific task tests. The most significant improvements lie in performance per task and token efficiency.

Compared to the previous V4 Flash version, the new 1017 update requires only one-quarter of the high-bandwidth memory. It also needs one-eighth of the solid-state drive storage for its KV cache. The model uses just 890 bytes per token entry. Deepseek backs these claims with a cost-per-task chart based on DeepSWE’s September 2 dataset. Both the standard and 1017 variants fall into the favorable top-left region of the chart. Even low-effort runs of the 1017 version complete tasks at a rate similar to the max settings of more expensive competitors. While DeepSWE is only one of many benchmarks, the results stand out.

Pricing remains a key advantage for the open-weight approach. At Deepseek’s off-peak rates, input costs 15 cents per million tokens. Output costs 60 cents per million tokens. Cache-hit input drops to 0.3 cents. Because the weights are open, the model can run on other platforms. Third-party providers have already undercut Deepseek’s API prices in the past. The company’s search feature costs 75 cents per 1,000 queries. This is lower than its rivals, with Parallel being the closest competitor at 1 dollar.

Artificial Analysis ranks the 1017 model second among large open-weight models on its intelligence index. It achieves a score of 45. This compares to 58 for the top-scoring Claude Opus 5.5 with fallback enabled. The model is among the fastest in its category. Only Xiaomi’s MiMo V2.6-Pro beats it on cost per task. These figures suggest the model is a strong contender for budget-conscious developers.

Reliability remains the primary drawback. On the Omniscience benchmark from Artificial Analysis, the model answered 43 percent of questions correctly. This is among the best results for open-weight models but lags behind proprietary systems. The model hallucinated on 67 percent of the questions it should have declined to answer. The benchmark index rewards correct answers and penalizes wrong ones, while non-answers do not count. The 1017 model scores 5 on this index. Kimi K3-max scores 20, and frontier models score 42 or higher.

This profile makes the model capable but best suited for specific jobs. Tasks with verifiable outputs work well. These include programming, OCR with review, and format conversion. An agent can check its work and retry if needed. Open-ended research, fact lookups, and drafting precise documents are better left to a frontier model or a human. In these cases, hallucinations are unlikely to be caught. The fast-and-loose behavior appears to be a trend among capable open-weight models. When Kimi K3 launched months ago, it was praised as cheap and effective. However, user experience ended quickly after short fact lookups produced invented information. Data from Artificial Analysis supports this observation. Open-weight models remain important for making AI agents more affordable. They also serve as a potential check on price rises from frontier labs.