🔍 Intelligence Synthesis JP/EN
Ethics · Regulation · Technology — Pharma Practice Notes

AI Tech Flash ── 2026-10-24

DeepSeek V4.1 Flash Update: Cheap, Smart, Error-Prone

A cheap, strong model that is reported to hallucinate often

What changes

According to Tom's Hardware, DeepSeek has upgraded its V4.1 Flash model to revision 1017. DeepSeek says the new version needs far less memory than its predecessor.

Tom's Hardware says the model suits tasks whose output can be verified, such as programming. It adds that open-ended research, fact look-ups and precise document drafting are better left to a frontier model.

Why it matters

Tom's Hardware says Chinese open-weight models broadly trade reliability for much lower running costs. Open-weight means the model's weights are published, so anyone can run it anywhere.

The outlet says such models make AI agents more accessible in dollar terms. An agent is an AI that carries out multi-step tasks on its own. The outlet adds that they could also help push down the cost of intelligence if frontier labs raise prices.

What it is

DeepSeek V4.1 Flash is an open-weight language model from DeepSeek. Tom's Hardware reports that it was just upgraded to revision 1017.

It is a Mixture-of-Experts (MoE) model, which runs only some of its specialist parts at a time to save computation. It has 552B (552 billion) total parameters, the model's internal adjustable values, and 16B of them are active during runs.

The model accepts text and images, and its context window is 1 million tokens. The context window is the amount of text it can handle at once.

Who and where

DeepSeek built the model. According to Tom's Hardware, it is open-weight, so you can run it anywhere.

DeepSeek also offers its own API. Its off-peak price per million tokens is $0.15 for input and $0.6 for output. On cache hits, input drops to $0.003. Tom's Hardware notes that third-party providers have undercut the DeepSeek API before.

DeepSeek also offers search at $0.75 per 1,000 queries.

Why it was built

DeepSeek's product page uses the slogan "wish-driven programming made real." Tom's Hardware reads this as a likely reference to vibe coding, where a user gives broad descriptions of what they want.

The outlet says long-running agentic tasks are where the cost of strong models matters most. The update targets that cost.

How it works

DeepSeek says the upgrade from V4 Flash to V4.1 Flash cut the KV cache needs sharply. The KV cache is the memory that holds intermediate results while the model processes a long input.

According to Tom's Hardware, V4.1 uses one-quarter of the HBM and one-eighth of the SSD storage of its predecessor. HBM is the fast memory attached to a GPU. It spends only 890 bytes per token entry.

DeepSeek also published a cost-per-task chart based on DeepSWE data. It says even low-effort runs of DSF-1017 come close to the "max" settings of pricey competitors.

Key points

  1. In DeepSeek's internal benchmarks, the model sits between GPT-5.6 Sol and Claude Opus 5, and beats Claude Opus 5 on some task tests.
  2. Artificial Analysis ranks DSF-1017 (max) second among large open-weight models on its intelligence index, with 45. Claude Opus 5.5 scores 58.
  3. On Artificial Analysis' Omniscience benchmark it answered 43% of questions correctly but hallucinated on 67% of non-answers. It scores 5 overall versus 42+ for frontier models.

Conclusion

Tom's Hardware reports that DeepSeek V4.1 Flash 1017 combines low cost with strong benchmark scores. The outlet names hallucination as its main drawback. It concludes the model fits tasks whose results can be verified.

Quotes and sources

“Deepseek V4.1 Flash (DSF) is a Mixture-of-Experts (MoE) model with 552B total parameters, 16B of which are active during runs.”
── Tom's Hardware, 10/23 12:00 UTC
“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”
── Tom's Hardware, 10/23 12:00 UTC
“the price at off-peak times per million tokens is just $0.15 for input and $0.6 for output.”
── Tom's Hardware, 10/23 12:00 UTC
“It answered questions correctly 43% of the time, which is among the best for open-weight models, but inferior to the proprietary competition.”
── Tom's Hardware, 10/23 12:00 UTC

Source articles

Written by AI (Claude) from the cited articles, then checked by Jev (TypeSafe's judgement model), a Gemini fact-check and a verbatim check of every quote.