Zhipu's GLM-5.3-FlashX hits 200 tokens/s — at 2.5x the price

Zhipu shipped the fast tier of its open model today and, unusually for a Chinese lab in 2026, raised the price to do it.

GLM-5.3-FlashX is live on the API at up to 200 tokens per second — roughly five times the throughput of GLM-5.3-Flash — and priced at about 2.5x what the base model costs. Zhipu says its 100,000-chip domestic inference cluster was saturated at launch under the weight of demand that never let up, and that the new tier follows fresh infrastructure investment and inference-side optimization rather than any change to the underlying model. It runs under its own model key, so developers pay the speed premium per call instead of inheriting it.

The base is the model we know. GLM-5.3-Flash shipped and open-sourced in August with 320 billion total parameters and 18 billion active in a mixture-of-experts layout — the first natively multimodal entry in the GLM-5 line, aimed at coding, agent work, visual programming and document handling. It also ran as the anonymous "Ox Alpha" on OpenRouter and OpenCode, where it became the highest-volume model on both before anyone knew whose it was. We covered the benchmark side of its launch in August — Z.ai's GLM-5.3-Flash tops benchmarks at one-tenth the price.

The interesting part is the pricing direction. For two years the Chinese open-model story was a race to zero, and a "Flash" badge meant cheap. FlashX keeps the same intelligence and the same architecture — hybrid linear-plus-sparse attention, fewer active parameters, native vision for checking its own front-end and Office output — and charges more for latency instead. That is a compute-scarcity admission: the constraint on serving agent traffic in China is no longer model quality, it is tokens per second on domestic silicon, and the lab is now selling the number customers can't get elsewhere. It also carves the market in two — price-sensitive batch work stays on Flash, latency-sensitive coding agents get billed a premium for waiting less.

Huawei used its Shanghai conference today to put a name on the same problem and a date on its export answer. At Huawei Connect 2026, Huawei Cloud CEO Peter Zhou said the company's enterprise agent platform AgentArts — already running inside more than 100 organizations, among them the Shenzhen Longgang district government, China Southern Power Grid and Kingsoft Office — will be commercially available outside China on December 30. Its open-source edition, openJiuwen, now carries more than 5,000 general and 1,000 industry-specific MCP assets, 50,000 GitHub stars and 3.29 million downloads, and Huawei is packaging partner commercial editions with Chinasoft International, iSoftStone and Beiming Software.

Huawei also published a seven-step "AI landing" method and the DIMAK engineering system spanning data, infrastructure, models, agents and knowledge, pitched at the failure mode of scattered agents with no process owner. Rotating chairman Eric Xu separately introduced the Peerium computing architecture, which links processors, NPUs, memory, storage and switches over an open bus protocol — the Atlas 950 supernode is its first commercial generation. The through-line is that Chinese infrastructure vendors are now selling deployment methodology, not just silicon.

What to watch: whether the speed premium holds once rivals push similar tiers, and whether AgentArts clears Western procurement review when it lands outside China.

Would you pay 2.5x for a faster model, or route the slow one and wait? Tell us in the comments.

Houmo moves compute inside memory — and A20 enters production

Beijing chipmaker Houmo used its partner conference this week to lay out two compute-in-memory routes at once — and to say the first chip built on the newer one is already in production. Houmo's A20, its first chip built on a 3D compute-in-memory architecture, has entered production. The part supports up to 16 chips fully interconnected, claims a 50% improvement in compute utilization over the previous generation, and covers models from 3B to 30B parameters, according to co-founder and product

Kling 4.0 goes live with 30-second native video at 10-bit HDR

Kuaishou's Kling AI has spent two years becoming the video model people actually finish projects with. Its new flagship, Kling 4.0, went fully live on October 15 — and this round is less about prettier frames than about the two things that make a generated shot usable in a real edit: length and control. Kling 4.0 generates up to 30 seconds in a single native pass, outputs 10-bit HDR at 4K, 1080p and 21:9, and takes up to 15 reference assets in one task. That budget splits across ten images, fi

Brain organoids classified Braille at 83% — in a trio, not alone

Two stories today: human brain tissue on a chip decoded a tactile signal, and the data-centre company behind some of the largest AI contracts in Europe turns out to be six months behind on its flagship site. Human forebrain organoids grown on a microelectrode array classified Braille letters with 61 percent accuracy on their own and 83 percent when the responses of three organoids were combined, in a peer-reviewed study from the University of Bristol. The team took event-based tactile readings

Suleyman says Microsoft's base model beat OpenAI's on a quarter compute

Microsoft finished pre-training its biggest in-house model yet — then used it to make a point about the company it bankrolls. Microsoft AI completed pre-training of MAI-Base-2, a mixture-of-experts foundation model with 80 billion active and 2.8 trillion total parameters, and says the recipe is 4.2–12.3x more pre-training FLOPs-efficient than DeepSeek-V4-Pro-Base. The run itself is the number worth keeping: 16,000 Nvidia GB200 GPUs for 38 days across August and September, at 21% average model

Read more

Houmo moves compute inside memory — and A20 enters production

Kling 4.0 goes live with 30-second native video at 10-bit HDR

Brain organoids classified Braille at 83% — in a trio, not alone

Suleyman says Microsoft's base model beat OpenAI's on a quarter compute