Topics / Open Models
Open Models
Open weights have become China’s distribution strategy. When a lab cannot out-spend the frontier on chips, it can still set the standard by making its stack the cheapest and most legal one to run anywhere. From DeepSeek V4 to a trillion-parameter model out of a food-delivery company, the open releases keep landing on domestic silicon.
The state of play
- Alibaba’s Qwen now leads the open-weight world outright. SCMP and Pandaily report it passed a billion downloads on Hugging Face this spring and accounts for more than half of every open-source model download on earth, running near 1.1 million downloads a day, having displaced Meta’s Llama, the default open model two years ago, which has fallen off the most-used rankings entirely; DeepSeek, MiniMax and Moonshot fill in most of the rest. Read the more-than-half and one-billion figures as that single circulating report’s, directional rather than audited. The lead is firm enough that Alibaba plans to start charging heavy users of its next open model, Qwen3.8-Max (Reuters), the kind of metering you do not attempt on something freely available elsewhere. In the same week Meta shipped an open model, Muse Glimmer, and said it will publish its flagship Muse Spark 1.2 weights, and Nvidia released Nemotron 3.5 Lightning alongside its training data, both re-entering an open-source market they had backed away from, explicitly to catch labs that are Chinese (CNBC; SCMP; Pandaily; Reuters, via CAD, “The Firm Flag”).
- DeepSeek open-sourced V4 under the MIT license in April 2026, with Huawei Ascend 950 as a primary deployment path.
- In May 2026 a research team completed what it says is the first third-party full-parameter post-training of a 1.6-trillion-parameter model on about 1,000 Huawei Ascend cards, with no instability.
- The bet: when a lab cannot out-spend the frontier on chips, open weights make its stack the cheapest and most legal one to run anywhere.
Best issues on this topic
- The Base Model
Harvey, the OpenAI-backed legal AI company, published its first post-trained model on 20 August. Harvey Tenet is a Kimi K3 base, post-trained with a rank-64 LoRA over the full network on a dataset of about 1,750 tasks. It completes almost twice as many held-out tasks on the Legal Agent Benchmark as base Kimi K3, and places second on that benchmark overall. Harvey ran ablations on candidate judge models against heavier frontier models and settled on Kimi 2.6 as the grader. Its M&A diligence work uses a GLM-5.2 orchestrator, taking rubric criteria pass rate from a 43.8 percent best baseline to 60.1 percent after self-distillation post-training.
- The Firm Flag
Alibaba’s Qwen now accounts for more than half of every open-source AI model download on earth and has passed a billion downloads on Hugging Face, displacing Meta’s Llama, per SCMP and Pandaily, though those share figures trace to a single report. The lead is firm enough that Alibaba plans to start charging heavy users of its next open model, and in the same week Meta and Nvidia both re-entered the open-weight market they had backed away from, Meta with Muse Glimmer and Nvidia with Nemotron 3.5 Lightning, explicitly to catch the Chinese labs now setting the open-source default.
- The Off-Ramp
DeepSeek finished shipping V4, and its day-0 domestic-chip support runs on Huawei Ascend (via Huawei’s CANN, its answer to Nvidia’s CUDA), Cambricon and Hygon. After the launch Alibaba, ByteDance and Tencent ordered hundreds of thousands of Ascend 950 processors.
- The Trade-Down
An IDC survey found 47 percent of large US firms now run a Chinese model in at least one use case, and at one point mid-July all five top models on OpenRouter were Chinese. Coinbase, DoorDash, Airbnb and Cursor are named adopters, switching to cut the bill as UBS puts Chinese API prices at 10 to 20 percent of the US alternative.
- The Giveaway
DeepSeek open-sourced an 85% inference speedup. OpenAI found the same trick the same fortnight and kept it secret. One bets on standards, the other on margin.
- The Week the Race Happened
DeepSeek open-sourced V4 under MIT with Huawei Ascend as a primary deployment path: 1.6 trillion parameters, million-token context, coding a hair behind the best Western models. The first sovereign stack.
- Closed in Code
A team completed what it says is the first third-party full-parameter post-training of a 1.6-trillion-parameter model on about 1,000 Huawei Ascend cards, with no instability. The penalty is efficiency, and it is shrinking.
These link to the full issues on the newsletter. New pieces on this topic go out in the daily first.
Common questions
Why does China open-source its AI models?
Open weights are a distribution and standards play. A lab that cannot out-spend the frontier on chips can still set the default by making its stack the cheapest and most legal one to run anywhere, especially on domestic silicon.
What license is DeepSeek V4 under?
DeepSeek released V4 under the permissive MIT license in April 2026, open-weighting the model with Huawei Ascend as a primary deployment path.
Which open-source AI model is the most downloaded in the world?
As of 2026 it is Alibaba’s Qwen. SCMP and Pandaily report it passed a billion downloads on Hugging Face and accounts for more than half of all open-source model downloads worldwide, at roughly 1.1 million a day, having displaced Meta’s Llama. Those share figures trace to a single circulating report and should be read as directional. DeepSeek, MiniMax and Moonshot account for much of the remainder.