China turns up the heat with open model blitz as US model makers panic

The AI arms race reached a fever pitch on Monday after Chinese e-commerce and cloud provider Alibaba called into question America’s technological lead with the launch of Qwen 3.8-Max, a 2.4 trillion-parameter model that goes toe-to-toe with the best models from Anthropic and OpenAI. The new model comes just days after the launch of DeepSeek V4 Flash 0731, which, according to independent benchmarks by Artificial Analysis, performs within a single point of OpenAI’s budget-friendly GPT-5.6 Luna while costing 40 percent less per task. What’s more, at just 284 billion parameters, it’s small enough to run on relatively modest enterprise servers and workstations. Chinese model devs like Moonshot, Alibaba, and DeepSeek are now attacking their American counterparts on both price and performance. The pincer movement comes as US model devs like Anthropic and OpenAI stoke fears over the origins and safety of China-made AI models. In a recent blog post, Anthropic CEO Dario Amodei insisted he’s not opposed to open models, just ones made in China, ones distilled from proprietary models, and ones that do not meet rigorous safety metrics. In other words, anything that actually competes with Anthropic's own models. The safety bit is particularly disingenuous, as the company’s fearmonger-in-chief has gone out of his way to stoke fears among US government officials. Proprietary models can be controlled, but open weights, once released in the wild, are impossible to claw back. But neither Amodei’s comments nor commitments from major American and European tech giants change the fact that China is providing the only meaningful competition in the open weights arena. China dominates here. Speaking on CNBC Monday, Clément Delangue, CEO of Hugging Face, the biggest and most influential model repo in the world, said as much. “They’re clearly dominating on open models right now, and I wouldn’t be surprised if they start dominating at the frontier either by the end of this year or next year at the rate of progress,” he told the financial TV news network. America’s most capable open weights model, Inkling, is just shy of a billion parameters, and still can’t keep up with DeepSeek’s finest. Models like Moonshot's Kimi K3 and Z.ai's GLM 5.2 are in an entirely different orbit. In fact, an enterprise's only credible alternatives to proprietary models and their dubious security policies are Chinese models. DeepSeek, Alibaba, Moonshot, MiniMax, and Z.ai aren’t going to pass that opportunity up. Alibaba lets its most powerful model loose on the world Of the Chinese model devs, Alibaba is arguably the most like its American competition. Its open-weight models are well regarded, and their range of sizes and permissive licensing have made them attractive for fine-tuning application-specific systems. However, like Google and OpenAI, its top models have been locked behind an API – until now. Seizing the moment, Alibaba is making its most capable model weights available for download for the first time with the launch of Qwen 3.8-Max. The blog post contains the usual array of vaguely intelligible bar charts showcasing how the model compares to OpenAI and Anthropic’s own models, as well as a slew of demos that would have made Billy Mays proud. If you want specifics, we recommend checking out the launch blog here. You don’t have to look that closely to get what Alibaba is selling: anything OpenAI and Anthropic can do, we can do as well, if not better, cheaper, and on the hardware you own. With that said, 2.4 trillion parameters is a rather big lift for most enterprises, likely requiring 48-64 Nvidia B200-class GPUs if deploying in a customer-facing capacity. For internal workloads, 8-16 B300 or AMD MI355X GPUs would be adequate. If that’s a little rich for your blood, Alibaba hasn’t forgotten its roots and will be releasing a 27-billion parameter version of the model alongside the Max variant. With benchmarks, model devs can usually find a collection of tests that paint their model in a positive light. And unsurprisingly, Artificial Analysis’ own Intelligence leaderboard tells a slightly different story than Alibaba’s, with Qwen 3.8-Max matching Anthropic’s less-capable, but still formidable, Claude Sonnet 5. Beyond the marketing and pitch demos, Qwen’s blog post is surprisingly short on detail. What we do know is it’s a multi-modal mixture of experts (MoE) model. That means of the 2.4 trillion total parameters, only 95 billion are actually used to generate tokens for any one request. We can also assume that the model uses the same hybrid Transformer+Mamba architecture as previous Qwen models in order to maintain performance across large contexts. Speaking of which, the model will support context windows up to 1 million tokens, though it’s still not clear whether that relies on techniques like rope scaling to extend it or not. The model is currently available via Alibaba’s API service QwenCloud for $2 per million input tokens, and $6 per million output tokens. Cached tokens are charged on a sliding scale with $0.25 charged per million implicit cached tokens, $0.17 per million explicit cache token reads, and $2.5 per million explicit cache tokens created. For comparison, Anthropic’s Claude Sonnet 5 will set you back $2/M input tokens, $0.20/M cache hits, and $10/M output tokens. And Sonnet 5 pricing is set to increase 50 percent starting September 1. Meanwhile OpenAI’s GPT 5.6 Luna, which falls just behind Qwen 3.8-Max and Sonnet 5 on the Artificial Analysis leaderboard, will run you $0.20/M input, $0.02/M cached input, $0.25/M cached write, and $1.20/M output tokens for short context lengths under 272,000 tokens and double that for jobs exceeding that context. The model weights are slated for release on popular model repos, including Hugging Face, starting next week. DeepSeek undercuts OpenAI with flashy new V4 refresh While Alibaba joins Kimi K3-maker Moonshot.AI’s assault on frontier models, DeepSeek has taken a very different tack with its latest open weights model: squeeze every ounce of performance from as few weights as possible. The result is DeepSeek V4-Flash-0731, a 284-billion parameter model that can fit into around 142 GB of GPU memory (at FP4). That means enterprises can easily run this model at scale on a single system. And despite its smaller stature, the Flash model actually outperforms the 1.6 trillion-parameter DeepSeek V4 Pro by nearly 14 percent on Artificial Analysis’ Intelligence leaderboard. With that said, the refinements made to DeepSeek V4 Flash will no doubt find their way into the Pro model before long. Like Qwen 3.8-Max, DeepSeek V4 Flash undercuts OpenAI and Anthropic on pricing – this time by a considerable margin. For API access, DeepSeek is currently asking $0.14/M input tokens, $0.0028/M cached tokens, and $0.28/M output tokens. However, in an agentic world filled with reasoning models, API pricing doesn’t paint a complete picture. A model may appear cheaper, but if it consumes twice as many tokens as another higher priced model, it may not actually be less expensive. Because of this, it's important to look at how efficiently the model solves real world problems. Alarmingly for the US LLM makers, according to Artificial Analysis, DeepSeek V4-Flash isn’t just cheaper per token; it is also incredibly efficient at its job. Compared to OpenAI’s GPT 5.6 Luna, which is among the most efficient and least expensive models in Sam Altman’s current lineup, DeepSeek’s latest model is a full 40 percent less expensive, with a cost to solve of just three cents versus five cents. That difference may feel small, but it’s worth remembering that for vibe coders consuming tens or hundreds of millions of tokens a day, that difference adds up quite quickly. One of the secrets to DeepSeek's efficiency is the integration of DSpark speculative decoding directly into the model weights. We’ve explored speculative decoding in the past, but in a nutshell, it involves using a smaller draft model to predict the outputs of a larger, more capable one. When it works, inference performance increases. When it guesses wrong, it falls back to the base model. More importantly, because the smaller model is only predicting the output of the larger one, it’s entirely lossless and therefore requires no compromise in terms of performance. Alibaba and others have implemented similar speculative decoding mechanisms, like multi-token-prediction (MTP), for this reason. With DSpark, DeepSeek claims it can extract 57–85% more per-user speed on the exact same hardware, which is an impressive claim, and one that at least in our testing rings true. Your own personal frontier model? DeepSeek V4 Flash is just small enough that running it at home is entirely possible if you’ve got some deep pockets, or a heck of a lot of memory lying around. Testing on a 128 GB DGX Spark, we were able to get the model running at a respectable 128,000 token context window using Unsloth’s IQ3-XXS quant in Llama.cpp. Three bits per weight gives us just enough room to pack the DSPARK draft model into memory, but it’s also a bit more compression than we typically recommend for homelab use. Because of this, Unsloth warns that the model’s outputs could show some signs of quality loss, but it does run. Unsloth has a full guide on how to get the model up and running, assuming you’ve got beefy enough hardware. We admit, the DGX Spark is not a cheap box. At $4,699, it’s squarely in workstation territory, as is the $3,999 Ryzen AI Halo we looked at early last month. But the fact that you can even contemplate running a model like DeepSeek V4-Flash at home is impressive in its own right. The original IBM PC kitted out with a monitor and diskette drive cost about $3,735 in 1981. Adjusting for inflation, that works out to about $13,700 in today’s money. If history repeats itself, the hardware necessary to run models like DeepSeek V4-Flash should become much more accessible within the next decade.®