DeepSeek V4.1 Flash revision 1017 cuts memory, not hallucination risk

By Sophie Lindqvist

What changed in revision 1017?

DeepSeek’s V4.1 Flash (DSF) revision 1017 uses one-quarter of V4 Flash’s high-bandwidth memory (HBM) and one-eighth of its SSD storage for the model’s KV cache at inference. That is the strongest efficiency claim in the release. DSF has 552 billion total parameters in a mixture-of-experts architecture. Only 16 billion parameters activate per run. It accepts text and images. Its context window reaches 1 million tokens.

DeepSeek says the revision improves performance per task and token efficiency. Its product slogan is “wish-driven programming made real.” That likely references “vibe coding.” The model targets open-ended work described broadly.

Reliability remains the unresolved claim. DSF is cheap in DeepSeek’s pricing and scores well in company tests. It is also hallucination-prone. Frontier laboratory models deliver stronger results at higher running cost. Chinese open-weight rivals often trade reliability for lower expense, particularly in long-running agentic tasks.

How much does the memory footprint fall?

DeepSeek reports 890 bytes of KV-cache storage per token entry. The memory figures are ratios, not absolute capacity. The release gives no total memory size, processor platform or production process. Those omissions prevent a full hardware-cost calculation. The published data do not permit comparison with a specific server configuration.

On a cost-per-task chart based on DeepSWE data from September 2, DSF and DSF-1017 occupy the preferred top-left region, while low-effort DSF-1017 approaches the task-completion rates of costly competitors’ maximum settings. DeepSWE supplies only one yardstick. The chart supports an efficiency claim for that dataset. It does not establish general reliability.

What does inference cost?

At DeepSeek’s off-peak API rates, 1 million input tokens cost $0.15, 1 million output tokens cost $0.60, and a cache hit reduces the input price to $0.003 per million tokens. Because V4.1 Flash is open-weight, operators can run it outside DeepSeek’s service. Third-party inference providers have undercut DeepSeek’s API in the past. Effective prices may therefore fall below the published figures.

DeepSeek also charges $0.75 per 1,000 search queries. Parallel, the closest cited rival, charges $1.

How do the benchmark results compare?

According to DeepSeek’s internal benchmarks, DSF sits between GPT-5.6 Sol and Claude Opus 5, and it beats Anthropic’s LLM in the company’s selected task tests. The company’s own ordering cannot establish performance on unseen workloads. It also provides no evidence of reliable abstention.

Artificial Analysis provides a third-party measure. It ranks DSF-1017 at maximum effort second among large open-weight models on its intelligence index. The model scores 45. Claude Opus 5.5, at maximum effort with fallback, scores 58 in the comparison.

Artificial Analysis also places DSF-1017 among the fastest models in its category. Only Xiaomi’s MiMo V2.6-Pro has a lower cost per task. Those rankings do not capture reliability.

What does the hallucination rate change?

Artificial Analysis’ Omniscience benchmark gives a less favorable result, placing DSF-1017 among the lowest models and showing 43% correct answers, a strong result among open-weight systems but one below proprietary rivals. The more serious number concerns missing knowledge: DSF-1017 hallucinates 67% of non-answers instead of reliably declining when it lacks supporting facts, a failure mode that can pass through an automated workflow without detection.

The benchmark’s main index rewards correct answers. It deducts points for wrong ones. Non-answers earn no credit. DSF-1017 receives 5 points. Kimi K3-max receives 20. Frontier models receive at least 42.

The benchmark does not make DSF-1017 incompetent. It makes task selection decisive. Plausible answers can conceal missing evidence. Verification must therefore sit outside the model for high-consequence work.

Which workloads can trust the output?

Programming, optical character recognition with review, and format conversions suit DSF-1017. Their outputs can be verified. An agent can test its work and retry. These workflows can absorb errors because the process catches them.

The model is less suitable for open-ended research, fact lookup, and the interpretation or drafting of highly precise documents, because hallucinated details can pass through unchecked when no independent verification step exists. A frontier model or human is safer.

Several months ago, Kimi K3 drew praise for effectiveness and low cost. Its promise was frontier-level intelligence at a fraction of the price. Bruno Ferreira’s trial was short-lived. Even brief, precise fact lookups returned fabricated information.

Ferreira wrote: “Fast and loose” seems to be a forming trend for highly competent open-weight models, too. He said Artificial Analysis data supported his experience. The anecdote is not proof. It does match the benchmark pattern.

Where does the cost advantage lead?

Open-weight systems still matter commercially. They can make agents affordable in absolute dollar terms. They can also restrain intelligence costs when frontier laboratories raise prices. If third-party providers continue undercutting DeepSeek’s listed rates, revision 1017 could widen that cost gap, but the 67% non-answer hallucination rate keeps the advantage concentrated in verified workflows.