Deepseek has released revision 1017 of its V4.1 Flash open-weight model, a point upgrade that cuts hardware requirements dramatically while keeping prices among the lowest in the industry. The trade-off is a familiar one for cheap models: reliability.

What You Need to Know

Deepseek V4.1 Flash is a Mixture-of-Experts model with 552B total parameters, 16B active, and a 1 million token context window. Revision 1017 uses one-quarter of the HBM and one-eighth of the SSD storage for its KV cache compared to its predecessor. At off-peak pricing, input costs $0.15 per million tokens and output $0.6, with cache hits at $0.003. The model performs well on verifiable tasks but struggles to recognize when it lacks supporting facts.

The Efficiency Gains Behind Revision 1017

The upgrade from V4 Flash to V4.1 Flash significantly reduced hardware demands. The new revision spends only 890 bytes per token entry for its KV cache, a fraction of what the previous version required.

Deepseek describes the model as "wish-driven programming made real," a nod to the colloquial "vibe coding" trend. The slogan suggests the model suits open-ended tasks with broad descriptions, though that framing sits uneasily with its reliability profile.

On Deepseek's internal benchmarks, the model slots between GPT-5.6 Sol and Claude Opus 5, even beating Anthropic's LLM in some task tests. Those results, however, come from the company's own evaluation rather than independent testing.

How DeepSWE and Artificial Analysis Rate the Model

DeepSWE's dataset from September 2 shows both DSF and DSF-1017 clustering in the desirable top-left area of a cost-per-task graph. Even low-effort runs of DSF-1017 offer task completion rates in the ballpark of max settings for pricier competitors.

Artificial Analysis ranks DSF-1017 (max) as only #2 on its intelligence index within the large open-weight class, with a score of 45. Claude Opus 5.5 (max with fallback) leads at 58. The model is one of the fastest in its category, and its cost-per-task is edged out only by Xiaomi's MiMo V2.6-Pro.

$0.15

Input per million tokens

$0.6

Output per million tokens

$0.003

Cache hit input cost

Deepseek also offers search functionality at $0.75 per 1,000 queries, undercutting competitors. The closest rival, Parallel, charges $1. Third-party providers have undercut Deepseek's API in the past, so those figures could drop further.

The Hallucination Problem That Undercuts the Value

The adage of "fast, cheap and reliable, pick two" applies squarely here. DSF-1017 is smart, but it does not know what it does not know.

Artificial Analysis' Omniscience benchmark ranks the model among the lowest on this measure. It answered questions correctly 43% of the time, among the best for open-weight models but inferior to proprietary competition. That would be acceptable if it skipped answers without a foundation. Instead, it hallucinated 67% of non-answers.

The benchmark's main index grants points for good answers and deducts them for wrong ones, with non-answers not counting. DSF-1017 gets 5 points, compared to Kimi K3-max's 20, with frontier models coming in at 42+.

  • Programming: Verifiable outputs let the agent check its work and retry.
  • OCR with review: Human oversight catches errors before they propagate.
  • Format conversions: Deterministic transformations are easy to validate.
  • Open-ended research: Hallucinations likely go unchecked without verification steps.

Tasks with verifiable outputs will be fine, as the agent can run verification steps against its work and try again. Open-ended research, fact lookup and drafting highly precise documents are better suited to a frontier model or a human reviewer.

Why This Matters

Deepseek's pricing strategy forces a reckoning for enterprises building agentic workflows at scale. A model that costs $0.15 per million input tokens makes long-running automation economically viable in ways that frontier models do not.

Teams that deploy DSF-1017 for verifiable tasks will capture real savings. Those that assign it open-ended research or fact-checking will inherit a hidden cost: the human review needed to catch fabricated answers.

The broader signal is that the open-weight tier is splitting into two camps. One competes on raw intelligence, the other on cost per completed task. Deepseek has chosen the second camp, and its hallucination rate is the price of admission.