Deepseek's Updated V4.1 Flash Cuts Memory Needs Sharply, but Hallucination Problem Persists

Deepseek has released a new revision of its open-weight V4.1 Flash model, labeled 1017. Though it reads like a minor point update, the change may represent a meaningful gain in efficiency. On the product page, the company uses the slogan "wish-driven programming made real," which appears to nod to so-called vibe coding and suggests the model suits open-ended tasks described in broad terms. Independent testing, however, indicates it is prone to hallucination.

The large language model market is broadly split in two. Closed commercial models from frontier labs are expensive to run but deliver excellent results, while Chinese open-weight rivals generally trade some reliability for much lower running costs, particularly in long-running agentic work.

V4.1 Flash is a Mixture-of-Experts model with 552 billion total parameters, 16 billion of which are active during a run. It is multimodal, accepting text and images, and supports a context window of 1 million tokens, now a common standard.

In Deepseek's own benchmarks, the model sits between GPT-5.6 Sol and Claude Opus 5, and it even beats Anthropic's model in task tests. The bigger gains are in performance per task and token efficiency. Compared with V4 Flash, V4.1 Flash needs only one-quarter of the HBM and one-eighth of the SSD storage for its KV cache, using just 890 bytes per token entry.

Deepseek backs the claim with a cost-per-task chart built on DeepSWE's September 2 dataset. Both the standard and 1017 variants land in the favorable top-left region, and even low-effort runs of the 1017 version complete tasks at a rate comparable to the "max" settings of costlier competitors. DeepSWE is only one of many benchmarks, but the result is impressive.

At Deepseek's off-peak pricing, input costs $0.15 per million tokens and output costs $0.6, with cache-hit input falling to $0.003. Because the weights are open, the model can be run elsewhere, and third-party providers have undercut Deepseek's API before. Deepseek's search feature costs $0.75 per 1,000 queries, below rivals, the closest being Parallel at $1.

Artificial Analysis ranks the 1017 model (max) second among large open-weight models on its intelligence index, with a score of 45, against 58 for the top-scoring Claude Opus 5.5 (max with fallback). The model is among the fastest in its category, and only Xiaomi's MiMo V2.6-Pro beats it on cost per task.

The drawback is reliability. On Artificial Analysis' Omniscience benchmark, the model answered 43% of questions correctly, among the best for open-weight models but behind proprietary systems. It also hallucinated on 67% of the questions it should have declined to answer. The benchmark's main index rewards correct answers and penalizes wrong ones, while non-answers do not count. The 1017 model scores 5, compared with 20 for Kimi K3-max and 42 or more for frontier models.

That makes the model capable but best suited to certain jobs. Tasks with verifiable outputs, such as programming, OCR with review and format conversion, should work well because an agent can check its work and retry. Open-ended research, fact lookups and drafting or interpreting precise documents are better left to a frontier model or a human, since hallucinations are unlikely to be caught.

The article's author notes that fast-and-loose behavior appears to be a trend among capable open-weight models. When Kimi K3 launched several months ago, it was praised as cheap and effective, but the author's own use ended quickly after even short fact lookups produced invented information, which Artificial Analysis data seems to support. Even so, open-weight models remain important for making AI agents more affordable and as a possible check on price rises from frontier labs.