Open Weights Just Stopped Being Second Best

August 5, 2026 The number that should worry every Western AI company right now isn't on any benchmark leaderboard. It's 29.55 trillion. That's how many tokens Chinese open-weight models processed on OpenRouter in a recent week. Eight of the platform's top ten models were built by Chinese labs — DeepSeek, Xiaomi, Tencent, Zhipu AI, MiniMax, StepFun, and Moonshot AI among them — and together those models accounted for roughly 87 percent of usage among the top ten, according to OpenRouter data. Volume by itself doesn't prove much. The capability story arrived in the space of seven weeks, between June 13 and August 3, when two Chinese labs released open-weight flagships that don't just undercut the US frontier on price — they beat it on specific leaderboards and match it on enough others to make "open weight" stop sounding like a consolation prize. This is the mid-2026 breakdown of the two models that matter: Alibaba's Qwen3.8-Max and Zhipu AI's GLM-5.2, what they do differently, and which one you should actually budget for.

Alibaba released Qwen3.8-Max on August 3, 2026, and immediately changed the math on what open-weight models can be. It's a sparse mixture-of-experts model with 2.4 trillion total parameters and roughly 95 billion activated per operation — about 25 times more total capacity than the compute it spends on any single token. The context window runs to one million tokens, and the model sees images and video, not just text.

The MIT case for GLM-5.2

Zhipu AI — operating internationally as Z.ai — released GLM-5.2 seven weeks earlier, on June 13, and chose a very different path to the same destination. MIT license. Weights on Hugging Face immediately. No regional restrictions. Text-only, deliberately, with a laser focus on coding and reasoning.

The verbosity tax

GLM-5.2 has a hidden cost that only shows up in production. MorphLLM's analysis found it produces roughly 43,000 output tokens per Index task, about 37,000 of them pure reasoning — versus 16,000 for GPT-5.5. The per-token price is genuinely cheap, but effective cost-per-task lands near $0.46, which erases much of the gap to the frontier.

The security tradeoff

Axios reported that hackers in Russian-language forums are openly discussing how easy GLM-5.2 is to jailbreak for hacking tasks. That's the flip side of an MIT license: unlike Claude or ChatGPT, an open-weight model can be downloaded, stripped of safety controls, fine-tuned against specific targets, and run with zero visibility to any provider. A security consultant told Axios: "An attacker can run it locally without safety guardrails, fine-tune it against their specific targets, and operate with zero visibility to any provider or defender." Graphistry, a cybersecurity firm, said GLM-5.2 is the first open-weight model it has tested that it would recommend for a "frontier-like" cybersecurity experience. CAISI assessments at NIST estimate its cyber capabilities are similar to Claude Opus 4.6, which shipped in February 2026.

Head to head, on the numbers

The pattern is consistent across every dataset. Qwen3.8-Max wins on raw capability, multimodality, and frontend output — the common community comparison is that it "sees more of the picture." GLM-5.2 wins on price, immediate availability, local deployment, and caution in engineering — the reputation is that it's "more precise."

That precision showed up in the DoNews controlled test, which put Qwen3.8-Max-Preview, GLM-5.2, and Kimi K3 in front of a real task: rebuilding a broken Next.js site running Payload CMS. GLM-5.2 demonstrated the deepest architectural understanding in the assessment round, identifying 14 CMS collections, Payload version locks, and database check mechanisms — but incorrectly flagged already-implemented full-text search as pending. Qwen finished the same analysis in under ten seconds with a precise capture of recent development activity. During bug diagnosis, Qwen directly started the CMS service and closed the loop fast; GLM reused an existing service when ports were occupied, which is mature engineering judgment, just slower. On precision modification, GLM fully preserved the site's auto-carousel, mouse drag, and loop logic with only minor frame skips, where Qwen deleted drag functionality and faked it with content copying. DoNews's overall verdict: Qwen delivered the best comprehensive performance. But GLM made fewer dangerous mistakes.

The wave underneath

Neither release is an outlier if you look at the broader season. BenchLM's July 2026 open-weight ranking has MiniMax M3 at the top (68.8) and GLM-5.1 close behind (66.9); the top ten on its Intelligence Index all come from China-based labs. Stanford's 2026 AI Index Report puts the best closed-to-open gap at roughly 3.3 percent as of March 2026 — still a gap, but down from the canyon of 2023-2024. Epoch AI estimates the average open-source lag at three to twelve months, with coding and math already at parity — or better — in specific tests.

The economics are doing the rest of the work. Forbes reported that Uber exhausted its entire 2026 AI budget in four months on Anthropic's Claude. Accenture reportedly urged staff to stop using AI for non-essential tasks. Chinese open-weight APIs run 60 to 90 percent cheaper than leading closed alternatives, and Morgan Stanley tracked US users routing more than 30 percent of their OpenRouter tokens weekly to Chinese open models from February through July, peaking at 46 percent. DoorDash and others now run a layered strategy: sensitive core tasks stay on US providers, routine analytics move to Chinese open models. Xiaomi's MiMo-V2.5 alone processed 10.5 trillion tokens in a single week, up 12 percent week-over-week, per OpenRouter's leaderboard.

The policy picture is just as turbulent. Nvidia CEO Jensen Huang has said publicly that US companies absolutely should be allowed to use Chinese AI models, and rallied 133 companies behind an open letter opposing restrictions. A coalition of nearly 200 Silicon Valley startups petitioned the White House to reject limits on open-weight access. Meanwhile, the White House has told top US AI companies that open-weight models from Chinese competitors will be exempt from mandatory safety testing under its new security framework — the scrutiny lands on closed frontier systems instead.

The market has voted in ways that are hard to ignore. Z.ai's annual recurring revenue reportedly hit $1 billion within five months of GLM-5.2's release — about three times faster than Anthropic's 15-month climb to the same mark — and its market value has reached HK$933.6 billion. Coinbase CEO Brian Armstrong confirmed the company now uses GLM 5.2 and Moonshot's Kimi 2.7 as its engineers' default tools. On Hacker News, a parallel conversation has started about whether hosted AI companies have a durable moat when users can switch models in a day — and about which jobs disappear when cheap agents can do the work.

The mid-2026 Best Of list

- Best overall open-weight model: Qwen3.8-Max. The PaperBench score alone justifies it; the multimodal support and #2 Vision Arena rank close the case.

- Best value: GLM-5.2. At $1.40/$4.40 per million tokens with an MIT license, it's the cheapest frontier-adjacent intelligence you can legally embed anywhere.

- Best for local deployment: GLM-5.2. 40 billion active parameters and 4-bit quantization make a single RTX 3090 a workable machine.

- Best for multimodal and frontend work: Qwen3.8-Max. Its Arena positions — #4 frontend code, #2 vision — exceed anything open-weight has done before.

- Best for math-heavy reasoning: GLM-5.2. AIME 2026 at 99.2 is the standout number across both releases.

- Best for research replication: Qwen3.8-Max. Alibaba's own numbers show a 28-point jump on PaperBench from the previous generation.

The honest bottom line is that they beat each other at different things. That's the actual headline of mid-2026. OpenRouter's own analysis concludes that the frontier labs do not, at this moment anyway, appear to be accelerating away from open-weight labs — and that costs will keep dropping for any fixed point of intelligence. Chris Clark of OpenRouter put it more plainly: "Open-weight models are having a moment in the sun as cost becomes a central focus of organizations with scaled AI usage."

This is not the moment open weight caught the frontier on every dimension. It's the moment "open weight" stopped meaning second best — and started meaning cheaper, more inspectable, and occasionally first.