DeepSeek: models, news and controversies

What is DeepSeek

DeepSeek is a Chinese AI company and the name of its language models, known for the open-weight DeepSeek-V models and the DeepSeek-R reasoning models. It drew global attention with V3 and R1, and in April 2026 launched V4, reportedly the largest open-weights model, running on Huawei chips. Founder Liang Wenfeng is its CEO. Since V4, DeepSeek has become the reference point for cheap models used in agentic work, where the AI reads files, writes code and retries after errors on its own.

Which DeepSeek models exist

The API currently offers two models: deepseek-flash (DeepSeek-V4.1-Flash) and deepseek-v4-pro (DeepSeek-V4-Pro-0813). Both have a 1 million token context and support thinking and non-thinking modes, according to DeepSeek’s documentation (as of 10 October 2026).

V4.1-Flash arrived on 10 September. It is a 552-billion-parameter multimodal model that cuts the key-value cache to a quarter of V4-Flash, to 890 bytes per token. At maximum effort it scored 90.6 percent on Terminal-Bench 2.1 and 74.2 percent on DeepSWE v1.1. The technical report adds a warning: the same checkpoint scored between 65.5 and 74.2 percent on DeepSWE depending only on which agent harness ran it, so gaps of a few points between models fall within noise.

V4 Pro left testing on 14 August as build V4-Pro-0813. Terminal Bench 2.1 went from 72.1 to 87.9 and DeepSWE from 12.8 to 62.7. Artificial Analysis rates it 53 on its Intelligence Index, level with GLM-5.2, behind Kimi K3 (60) and Claude Opus 5 (63). Reports conflict here. An 11 September report says that from 14 September all traffic to the deepseek-v4-pro endpoint is rerouted to V4.1-Flash. DeepSeek’s documentation as of 10 October still lists deepseek-v4-pro as a separate model, V4-Pro-0813, with its own price list. Check which model answers before relying on either.

Vision model. On 24 August DeepSeek released the experimental V4-Flash-Vision-Exp with image understanding. It accepts JPEG, PNG, GIF and WebP, at most 384 tokens per image and up to 600 images per request. The documentation now says the vision-exp names are retired and deepseek-flash handles images.

Is DeepSeek better than ChatGPT and Claude

Not at the top end, but close, and far cheaper. V4 Flash 0731 scored 50 on the Artificial Analysis Intelligence Index, one point behind OpenAI’s GPT-5.6 Luna, at about 60 percent lower cost per task. In another comparison V4 Flash Max scored 50 against 56 for Claude Opus Max, but running the evaluation cost $72.02 versus $3,752.55. The best DeepSeek model, V4 Pro, trailed the leader, Claude Opus 5, by ten points in the 14 August ranking.

The market agrees. DeepSeek’s share of tokens on OpenRouter roughly doubled from 9 percent in January to 18 percent in June, and since mid-May V4 has been the top model there. In a single day V4 Flash consumed 8 trillion tokens on the OpenCode coding tool. Microsoft, meanwhile, considered a self-hosted DeepSeek V4 as a cheaper optional model for Copilot Cowork on Azure. See Microsoft Copilot.

DeepSeek also runs inside other vendors’ tools. Since 24 September Claude Code can be pointed at DeepSeek without a proxy: set the base URL to api.deepseek.com/anthropic, use a DeepSeek key and the model deepseek-flash. Codex CLI, OpenCode and Aider connect similarly. Background: Claude Code and OpenAI Codex. DeepSeek also has its own agent, Harness v0.1 under the MIT license, built on the Cordis plugin system.

How much does DeepSeek cost and is it free

The DeepSeek API is paid per token, while some models can be downloaded free under the MIT license. The official homepage does not state a price for the app or web chat, so I leave that out. API prices per DeepSeek’s official pricing page (as of 10 October 2026, USD per million tokens):

Peak hours are 01:00-04:00 and 06:00-10:00 UTC, Monday to Friday, excluding Chinese public holidays. Everything else is off-peak and costs half. The page does not mention a free API tier.

Prices have gone up. In July the cheapest V4 Flash endpoint cost $0.09 per million input tokens and $0.18 per million output. In August DeepSeek raised V4 Pro rates from 16 August: off-peak input from $0.435 to $0.66, output from $0.87 to $1.98, with cache hits rising the most. For agents that reread the same files, that is the most expensive part of the change.

Is DeepSeek open source and can you run it locally

Largely yes, with a caveat about the newest weights. The V4 Flash 0731 weights are MIT-licensed on Hugging Face (284 billion parameters, 13 billion active, 1 million token context). When V4-Pro-0813 launched, DeepSeek did not publish the new build’s weights, and the April version stayed on Hugging Face. As of 10 October DeepSeek’s official Hugging Face account lists V4.1-Flash and V4-Pro-0813 among its models, but shows no license there, so I do not state one. The Harness agent and the inference-speedup framework DSpark, up to 85 percent faster, are also MIT-licensed. Risks of open weights are covered in open-weight model security.

Local use is hard. V4 Flash weights at 16-bit precision need 568 GB of memory. The DwarfStar project, led by the creator of Redis, aims to run the V4 family on consumer hardware: it uses 2-bit quantization for less critical parts and 4-bit for key ones, streams from SSD and can spread inference across devices. Ollama also runs models locally without the API.

Is DeepSeek safe

Its safeguards are weak, and the model has been used in real attacks. The UK AI Security Institute (AISI) found the open models’ safeguards largely ineffective: DeepSeek V4-Pro sometimes refused reverse-engineering tasks, but trying again bypassed the refusal. AISI also rated V4-Pro’s cyber skill at the level of Opus 4.5 from November 2025, with a solved task costing about 28 cents against $12.50 for Opus 4.5.

Documented misuse:

• Autonomous attacks. Unit 42 reported that a Chinese-speaking actor in Zhuhai first tried Claude and OpenAI models, which refused, then wired DeepSeek into the Hermes Agent framework as an autonomous operator. Unit 42 called it the first confirmed proof that provider safety controls have operational value. See autonomous AI agent cyberattacks.
• Ransomware. Check Point found that DeepSeek produced a working ransomware technique from a broad prompt, targeting Android photos through the browser. V4 refused a direct request but complied once explicit terms were removed. See ransomware.
• Government intrusions. Suspected China-linked operators used Claude Code together with DeepSeek-v4-pro to breach systems in four countries, with Claude Code running execution and phishing pages and DeepSeek generating scripts. See state-backed hackers using AI.
• Malware with a model panel. Cisco Talos described ClosedQuorum, an implant that lets AI models vote on its next move; DeepSeek breaks ties. See AI-powered malware.

Political lean is a separate matter. A Washington Post investigation found that DeepSeek V4 Pro gave only left-leaning arguments in 70 percent of answers, against 80 percent for GPT-5.5 and 43 percent for Claude Opus 4.8.

Who owns DeepSeek and what about chips

DeepSeek is led by founder and CEO Liang Wenfeng and took no outside money until June 2026. On 17 June it closed its first outside funding round: over 50 billion yuan, about $7.4 billion, at a valuation above $50 billion. Investors, including Tencent and CATL, put money into a limited partnership managed by the CEO, with no voting rights and a five-year lock-up. China’s state-backed AI investment fund invested directly and kept voting rights. In August the API price rise coincided with capital raising and preparation for an IPO.

The second thread is independence from Nvidia. On 1 October DeepSeek and Huawei released open-source tools for Ascend chips: compute and communication libraries plus TileLang, which DeepSeek says is simpler to program than Nvidia’s CUDA. They also optimized a supernode of 128 Ascend 950 chips. Context: AI chips and AI export controls.

What it means for you

• Compare cost per task, not price per token. The gap to closed models runs to tens of times, but API prices already rose in August and cache costs more than before the May cut.
• Treat DeepSeek’s safeguards as thin. If you run it in an agent with real access, add your own limits, because AISI bypassed refusals by simply retrying.
• Test in your own setup. The same checkpoint scored 65.5 to 74.2 percent on DeepSWE depending on the harness, so vendor tables are a hint, not a guarantee.
• When switching endpoints, check which model actually answers: the deepseek-v4-pro and vision-exp names have changed behavior.

Still open: whether the V4.1-Flash weights will carry the same MIT license as V4 Flash 0731, whether DeepSeek keeps its prices with new investors on board, and whether Ascend chips catch up with Nvidia.