On February 6, an anonymous model called Pony Alpha appeared on OpenRouter — free, 200K context, optimized for agentic workflows. Within 48 hours it topped the platform’s popularity rankings, and developer @gosrum reported it achieved a perfect ts-bench score matching Claude Opus 4.6. On February 11, Zhipu AI confirmed Pony Alpha was GLM-5: a 745-billion-parameter Mixture-of-Experts model trained entirely on domestic Chinese chips.
But the headline parameter count is marketing. GLM-5 activates only 44 billion parameters per token — roughly 5.9% of its full weight — on hardware that runs at 60% of NVIDIA’s best. That efficiency gap is the real story.
Zhipu, which hit 84.9% on LiveCodeBench V6 with December’s GLM-4.7, IPO’d on the Hong Kong Stock Exchange on January 8 at HK$116.20 per share. The GLM-5 reveal sent its stock up more than 60% in two trading sessions, pushing market cap past HK$150 billion. The model itself, though, tells a more interesting story than the ticker.
The 44 billion parameter reality behind GLM-5
GLM-5 uses a Mixture-of-Experts architecture with 256 total experts, of which only 8 fire per token. That yields approximately 44 billion active parameters per inference — a compute profile closer to a model between Llama 3 70B and a smaller dense frontier system than to the 745B headline number. As Reuters reported, the model approaches Claude Opus 4.5 in coding benchmarks and surpasses Gemini 3 Pro on certain tests.
The architecture borrows DeepSeek Sparse Attention (DSA) for efficient long-context processing across its full 200,000-token context window — the same mechanism DeepSeek itself uses and, as of early February 2026, scaled to over one million tokens. First-token latency sits under one second for short prompts, with sustained throughput of 30-60 tokens per second during streaming. The model represents roughly a twofold parameter increase over GLM-4.7, but the active compute footprint barely changed. None of this required a single NVIDIA chip.
Zhipu trained GLM-5 on Huawei Ascend chips — including the Atlas 800T A2 and likely the Ascend 910C — using Huawei’s MindSpore framework. The company also tapped processors from Moore Threads, Cambricon, and Kunlunxin. It’s a fully domestic Chinese hardware stack. As Computerworld reported, Zhipu claims this marks “the first time a state-of-the-art multimodal model completed its full training cycle on Chinese-made chips.”
Reframe that: 44 billion active parameters, running on chips that DeepSeek researchers concluded delivers roughly 60% of H100 effective performance, produced frontier-competitive results. This isn’t a scale achievement. It’s an engineering efficiency achievement — and that distinction matters for everyone betting that hardware restrictions would prevent exactly this outcome.
Pony Alpha: the stealth launch that validated GLM-5
OpenRouter introduced Pony Alpha on February 6 as a “stealth model” — no attribution, no marketing, just a free API endpoint. Developers immediately started testing. Japanese developer @gosrum ran it through ts-bench and reported a perfect score, matching Claude Opus 4.6, Opus 4.5, and Kimi K2.5. Chinese developer @hylarucoder spent half a day testing it against a multi-thousand-line Rust and TypeScript coding project and concluded its coding ability “clearly exceeds Sonnet 4.5 by a lot.”
The speculation was immediate and wildly varied: DeepSeek V4? Anthropic Sonnet 5? The guessing game mattered less than the market’s bet. South China Morning Post reported that Zhipu’s stock surged 36% on February 9 — before any official confirmation. Chinese financial media was connecting the dots faster than the developer community.
When Zhipu confirmed the identity on February 11, the stealth strategy had already accomplished its goal: organic developer validation without a dollar of paid marketing. It’s the same playbook DeepSeek ran with R1 — let the model speak first, attach the name later.
GLM-5 benchmarks: what we know and what’s missing
The confirmed numbers paint an incomplete but compelling picture. GLM-5 approaches Claude Opus 4.5 in coding benchmarks and surpasses Gemini 3 Pro on select tests. The reported ts-bench perfect score would match Opus 4.6 and 4.5 exactly, though this result comes from independent developer testing rather than the official leaderboard. For context, its predecessor GLM-4.7 hit 84.9% on LiveCodeBench V6 and 73.8% on SWE-bench Verified — GLM-5 is expected to exceed both substantially.
What’s missing is equally telling. As of launch day on February 11, independent SWE-bench Verified, AIME, and HLE scores haven’t been published. Self-reported benchmarks and third-party evaluations will need reconciliation once more labs run their own tests. The anonymous free-testing conditions of the Pony Alpha launch — zero cost, no rate limiting — may not reflect production API performance at scale.
That caveat matters, but it cuts both ways. Developers testing Pony Alpha weren’t running sanitized benchmarks — they were building real projects. @hylarucoder’s half-day Rust and TypeScript session is the kind of sustained, practical evaluation that synthetic benchmarks can’t replicate. The model earned its reputation before anyone knew its name.
The export control counterexample
In January 2025, the US Commerce Department placed Zhipu AI on the Entity List, designating it as acting “contrary to US national security.” The designation effectively cut Zhipu off from US-controlled technology — including advanced NVIDIA chips and cloud services — by requiring export licenses that are presumptively denied. Thirteen months later, Zhipu launched a model that matches frontier Western systems on coding benchmarks.
The Huawei Ascend 910C delivers roughly 800 TFLOPS at FP16 — about 80% of H100’s raw computing power. In practice, DeepSeek researchers measured effective performance at closer to 60%. The gap is real, measurable, and documented. It was also clearly insufficient to prevent frontier model development.
Every Washington policy hawk who argued chip export controls would maintain a two-to-three-year US AI lead now faces a benchmarkable counterexample. The debate will split predictably: one camp will push to tighten restrictions further, targeting TSMC, Samsung, and SK Hynix components in the Ascend supply chain. The other will argue the controls accelerated China’s domestic chip ecosystem — that restriction bred self-sufficiency. Both camps can find evidence in GLM-5’s spec sheet.
An important caveat deserves its own paragraph: the Ascend 910C still relies on TSMC manufacturing, Samsung memory, and SK Hynix parts. Full supply chain independence — chips designed, fabricated, and assembled entirely within China — remains incomplete. GLM-5 is the proof of concept for a post-NVIDIA training stack. The chips underneath it aren’t yet proof of a post-TSMC fabrication stack.
Open source with an asterisk
Zhipu plans to release GLM-5 under an MIT license in Q1 2026, following the precedent set by GLM-4.7 on Hugging Face. The pricing story is equally aggressive: GLM-4.7 standard input runs $0.60 per million tokens — less than half of GPT-5’s $1.25 — with cached input dropping to $0.11 per million. During the Pony Alpha stealth launch, the model was free — zero dollars per million tokens.
Here’s the asterisk. China’s export control catalog restricts certain AI-related technologies from leaving the country, and regulators have broad discretion over what qualifies. On the other side, US export controls now cover advanced AI model weights under ECCN 4E091. An MIT license promises unrestricted commercial use, but both governments can limit where weights actually flow.
If a Western developer downloads GLM-5 weights and runs them on AWS us-east-1, the legal picture is murky on both sides of the Pacific. The “open source” label is doing heavy marketing work while the regulatory reality points toward geographic friction.
This tension isn’t unique to GLM-5. DeepSeek — whose context window quietly expanded 10x to over one million tokens around the same day as the GLM-5 launch, per South China Morning Post — faces the same legal overhang. Both companies adopted DeepSeek Sparse Attention. Both are scaling aggressively. And both are navigating overlapping export control regimes on each side of the Pacific that increasingly treat model weights as controlled technology.
What GLM-5 actually proves
If a chip running at 60% of H100 performance produces a model that matches Claude Opus on coding benchmarks, what exactly is the performance threshold below which export controls actually work — and does that threshold exist?
GLM-5 isn’t the model that proved Chinese AI caught up. It’s the model that proved the catching up happened while Washington was still debating whether restrictions were strict enough. The 745B parameter headline will dominate the news cycle. The 44B active parameter number — doing frontier-class work on hardware nobody in Washington thought was sufficient — will dominate the next round of policy memos.
DeepSeek V4, slated for mid-February 2026 with one trillion parameters and full one-million-token context targeting 80%+ on SWE-bench, will determine whether GLM-5 is an outlier or the new Chinese AI baseline. That answer is days away as of February 11 — and it will reshape the entire debate.
Get the Daily Pulse
Sharp analysis on what's actually moving in AI. No hype, no filler, no weekly digest.
