Alibaba and DeepSeek have unleashed open-weight models that match frontier US performance at a fraction of the price—here's what that means for your infrastructure and wallet.
Source material: theregister.com
Alibaba's Qwen team has released the weights of its most capable model, Qwen 3.8-Max, for the first time, ending a period where its top models were API-only. The 2.4-trillion-parameter mixture-of-experts model activates only 95 billion parameters per request, keeping inference feasible. It matches Anthropic's Claude Sonnet 5 on the Artificial Analysis Intelligence leaderboard while costing $2 per million input tokens and $6 per million output tokens, with cached reads at $0.17 per million. Weights will be downloadable from Hugging Face next week. For enterprises, this means you can self-host a frontier-level model without sending data to a US API provider. Hardware demands are steep: internal workloads need 8–16 Nvidia B300 or AMD MI355X GPUs; customer-facing deployment requires 48–64 B200-class GPUs. A 27-billion-parameter version accompanies it for lighter users.
DeepSeek V4 Flash needs only 142 GB of GPU memory in FP4, so a 128 GB DGX Spark can run it with Unsloth’s IQ3-XXS quant in Llama.cpp at a 128,000-token context window; Unsloth warns the heavy compression can degrade quality. The $4,699 DGX Spark is workstation territory, as is the $3,999 Ryzen AI Halo. Qwen 3.8-Max is a different scale: it is a 2.4-trillion-parameter mixture-of-experts model with 95 billion active parameters, but deployment still calls for 8–16 Nvidia B300 or AMD MI355X GPUs for internal workloads and 48–64 Nvidia B200-class GPUs for customer-facing service. Alibaba will also release a 27-billion-parameter Qwen variant for more modest hardware. For perspective, a fully outfitted IBM PC cost about $3,735 in 1981, or roughly $13,700 today; the trajectory of commodity hardware suggests today’s workstation-class AI boxes should become far more accessible within a decade.
Hugging Face CEO Clément Delangue told CNBC that China is "clearly dominating on open models right now" and that he would not be surprised if Chinese labs start dominating at the frontier "by the end of this year or next year" at the current rate of progress. The open-weights roster is nearly all Chinese: Alibaba, DeepSeek, Moonshot’s Kimi K3, Z.ai’s GLM 5.2, and MiniMax all publish downloadable weights, while America’s most capable open model, Inkling, has under a billion parameters and still cannot keep up with DeepSeek’s best. Anthropic’s Dario Amodei says he opposes only Chinese-made open models, distilled models, and models without his preferred safety metrics, but that stance does not change the practical choice for enterprises: at the performance tier where self-hosting is viable, the only open weights with permissive licensing and frontier-class results come from China. Open weights cannot be pulled back after release, and Alibaba’s flagship is already scheduled for Hugging Face next week.
Speculative decoding is an inference accelerator: a small draft model predicts the large model’s next tokens, and when the guesses are right, the large model validates several tokens in one pass; when a guess is wrong, it recalculates, so the final output is identical. DeepSeek bakes the DSpark draft directly into V4 Flash’s weights rather than running a separate component, which lowers overhead and makes the speedup almost free. Alibaba uses a similar mechanism, multi-token prediction, in its Qwen models. DeepSeek claims DSpark gives 57–85% more per-user throughput on the exact same hardware, and hands-on testing found that claim believable. For someone choosing a model, the key consequence is cost-per-task: a model that spends fewer tokens to reach the same answer can be economically cheaper even before comparing listed rates. That is how DeepSeek undercuts OpenAI’s GPT-5.6 Luna on real workloads while staying within a point on benchmarks, and why an API price sheet by itself can mislead.
API pricing varies widely across the leading models. Alibaba's Qwen Max charges $2 per million input tokens and $6 per million output, with a sliding cache scale: $0.25 per million implicit cached tokens, $0.17 per million explicit cache reads, and $2.5 per million explicit cache tokens created. Anthropic's Claude Sonnet 5 asks $2 per million input, $10 per million output, and $0.20 per million cache hits, with output price rising 50% on September 1. OpenAI's GPT-5.6 Luna costs $0.20 per million input, $0.02 per million cached input, $0.25 per million cache writes, and $1.20 per million output for contexts under 272,000 tokens, doubling for longer ones. DeepSeek undercuts them all: $0.14 per million input, $0.0028 per million cached tokens, and $0.28 per million output. The gap matters most for agentic workloads, which hammer on cached context; DeepSeek's repeat-prefix price is a fraction of the others, though its input and output rates are also the lowest.
Anthropic CEO Dario Amodei says he is not opposed to open models, only ones made in China, ones distilled from proprietary models, and ones that do not meet his preferred safety metrics—which, in practice, would exclude the open models that actually compete with Anthropic. He has stoked those concerns among US government officials. The technical reality cuts both ways: proprietary models can be controlled, but open weights, once released, are impossible to claw back, so any harmful behavior must be handled by users. At the same time, self-hosting a Chinese model keeps proprietary data out of US API pipelines and gives the deployer full control over how the model is run and fine-tuned. Alibaba’s flagship is slated for public repos and DeepSeek’s latest is already downloadable; no vendor can revoke them after download. The practical conclusion for an enterprise is to do its own evaluation before deployment—run internal tests on sensitive tasks, monitor outputs, and treat the open-weights model as the deployer’s responsibility.
DeepSeek V4 Flash's efficiency comes from DSpark, a speculative decoding mechanism baked directly into the model weights. Instead of generating every token with the full 284-billion-parameter model, a smaller draft model predicts likely outputs; when a guess matches, the larger model skips ahead, and when it errs, the system falls back to the base model. This is lossless—no quality is sacrificed because the draft only proposes, the base model remains the arbiter. DeepSeek claims the technique yields 57–85% faster per-user inference on identical hardware. The practical benefit: a 284-billion-parameter model that fits in roughly 142 GB of GPU memory at FP4 can serve enterprise workloads without a cluster. On a 128 GB DGX Spark, the model runs with a 128,000-token context window using Unsloth's IQ3-XXS quant, though that aggressive three-bit compression may introduce slight quality loss. For homelab enthusiasts, the $4,699 DGX Spark or $3,999 Ryzen AI Halo bring such capability within reach—a far cry from the $13,700 (inflation-adjusted) for an original IBM PC, suggesting that local frontier-scale inference will become increasingly accessible over the next decade.
Where this came from. This breakdown is based on source material published at theregister.com. Images above are used with the credits shown beneath each one.
