Deepseek just upgraded its V4.1 Flash open-weight model to revision 1017. What looks like a mere point revision may actually be a significant step forward in model efficiency. On the product page, Deepseek uses the slogan "wish-driven programming made real," a likely reference to the colloquial "vibe coding," implying the model is a good fit for open-ended tasks with broad descriptions. However, while the model is indeed cheap and scores well, it can be hallucination-prone.
LLMs are roughly divided into two categories: closed commercial efforts from frontier labs, whose models are expensive to run but deliver excellent results, and Chinese open-weight competitors that broadly exchange reliability for a much cheaper running cost, particularly for long-running agentic tasks.
Deepseek V4.1 Flash (DSF) is a Mixture-of-Experts (MoE) model with 552B total parameters, 16B of which are active during runs. The model is multimodal, accepts both text and images as inputs, and its context window length measures the now-standard 1 million tokens.
According to Deepseek's internal benchmarks, the model slots in neatly between GPT-5.6 Sol and Claude Opus 5, even beating Anthropic's LLM in task tests. That's already a strong showing, but this latest revision of DSF shines in the performance-per-task base and token efficiency. The upgrade from V4 Flash to V4.1 Flash significantly slashed hardware requirements, with V4.1 using only one-quarter of the HBM and one-eighth of SSD storage for its KV cache compared to its predecessor, spending only 890 bytes per token entry.
The company illustrates the point with a cost-per-task graph based on DeepSWE's dataset from September 2, adding its own data points for both DSF and DSF-1017. Both variants cluster in the desirable top-left area, and even low-effort runs of DSF-1017 offer a task completion rate in the ballpark of "max" settings for pricey competitors. It's worth noting that DeepSWE is only one of many benchmarks, but this is nonetheless an impressive result.
Even at Deepseek's own pricing (remember, this is an open-weight model you can run anywhere), the price at off-peak times per million tokens is just $0.15 for input and $0.6 for output. On cache hits, the input token cost drops to a mere $0.003. It's worth keeping in mind that third-party providers have undercut Deepseek API in the past, so those figures could go lower if you shop around inference providers. Deepseek also offers search functionality at $0.75 per 1,000 queries, an amount that undercuts competitors, with the closest one being Parallel at $1.
DeepSWE isn't the only benchmark around, either. Artificial Analysis ranks DSF-1017 (max) as only #2 on its intelligence index within the large open-weight class, with a score of 45. For reference, the most seemingly intelligent model, Claude Opus 5.5 (max with fallback), rings in at 58. It's one of the fastest models in its category, too, and its cost-per-task is edged out only by Xiaomi's MiMo V2.6-Pro.
This all sounds too good to be true, but everyone knows the adage of "fast, cheap, and reliable, pick two." So what's the catch? In one word: hallucination. Just because a model is smart doesn't mean it's reliable. Broadly speaking, users prefer models that don't answer when they lack supporting facts, and DSF-1017 doesn't seem very good at knowing what it doesn't know.
Artificial Analysis' Omniscience benchmark is quite clear on this matter, ranking the model among the lowest. It answered questions correctly 43% of the time, which is among the best for open-weight models, but inferior to the proprietary competition. That would be perfectly acceptable if it skipped answers without a foundation, but it hallucinated 67% of non-answers.
The benchmark's main index grants points for good answers and deducts them for wrong ones, and non-answers don't count. DSF-1017 gets 5 points, compared to Kimi K3-max's 20, with the frontier models coming in at 42+. Ultimately, this means DFS-1017 is by no means incompetent — much the opposite —but one should be careful about which tasks to assign to the model.
Tasks with verifiable outputs like programming, OCR with review, and format conversions will be fine, as the agent can run verification steps against its work and try again. However, open-ended research, looking up facts, and interpreting or drafting highly precise documents are tasks better suited to a frontier model — or, you know, a human — as hallucinations likely won't be properly reined in and results won't be very trustworthy.
"Fast and loose" seems to be a forming trend for highly competent open-weight models, too. Several months ago, when Kimi K3 was released, everyone raved about how effective and cheap it was, and how it proposed to offer frontier-level intelligence for a fraction of the price. My personal experience with Kimi was short-lived, as even short but precise fact look-ups came up with hallucinated information, and Artificial Analysis' data seems to back this observation up.
Nevertheless, open-weight models remain important both to make AI agents more accessible in absolute dollar terms and potentially as a tool to drive down the cost of intelligence when facing down potential price rises from frontier labs.