Chinese Open Weight Models Top the Leaderboards. What Can You Actually Run?

Chinese Labs Own the Open-Weight Leaderboards. That Doesn’t Mean You Can Run Them.

Open-weight model rankings in mid-2026 read like a list of Chinese labs: Z.ai’s GLM, Alibaba’s Qwen, DeepSeek, Moonshot’s Kimi.

And the biggest release of the year has now landed. Moonshot published Kimi K3’s full weights on July 27, 2026, making it the largest open-weight model released to date.

“Open weights” is doing a lot of work in that sentence, though. It means the file is free to download. It does not mean the file fits on your graphics card.

The gap is enormous, and almost nobody quantifies it. So we did. Every number below comes from our own calculator, which computes what a model needs when it’s held entirely in VRAM. The spread within a single category, “Chinese open-weight models”, runs from 7 GB to over a terabyte. Both ends are genuinely open. Only one end is genuinely yours.

The Frontier Tier: Open, and Completely Out of Reach

These are the models topping the rankings. At Q4 quantization, the common GGUF default, here’s what full residency costs:

Model	Total params	VRAM needed	Smallest pool that fits
DeepSeek-V4-Pro	1.6T / 49B active	~1,116 GB	7× B200
Kimi K2.5	1T / 32B active	~698 GB	5× B200
GLM-5.2	753B	~526 GB	4× B200
DeepSeek-R1	671B / 37B active	~469 GB	3× B200
Qwen3 Coder 480B	480B / 35B active	~337 GB	2× B200

That’s a datacenter rack, not a desk. GLM-5.2 is MIT licensed, carries no regional restrictions, and is genuinely free to download. It needs roughly half a terabyte of VRAM to run the way it was designed to run. The license is open. The hardware bill is not.

This is the part the leaderboard coverage skips, and it’s why “the open-weight gap has closed” is only half true. The capability gap closed. The access gap moved somewhere else: into your power supply.

The clever part: these models are lean where it counts

There’s a genuine engineering story inside those numbers, and it’s worth understanding because it’s why the next tier exists at all.

Look at the KV cache, the memory that grows with your conversation length. GLM-5.2 is a 753B-parameter model, and its KV cache at 8K context is under 1 GB. DeepSeek-V4-Pro’s is 49 MB. For comparison, Qwen3 32B, a model twenty times smaller than GLM-5.2, uses 2.1 GB.

The big models are more memory-efficient per token of context, not less. Three techniques do it:

Compressed (latent) KV cache. GLM-5.2, DeepSeek, and the Kimi K2 line all store one compressed 576-dimension vector per layer instead of full per-head keys and values. We read this straight from GLM-5.2’s own config file: kv_lora_rank: 512 plus a 64-dimension rotary component.
Sparse attention with shared indexers. GLM-5.2’s config lists 78 layers, of which only 21 store their own attention indexer. The other 57 reuse them. Z.ai calls this IndexShare. It’s the difference between a manageable cache and an unmanageable one at million-token context.
Mixture-of-Experts. GLM-5.2 activates 8 of 256 experts per token. DeepSeek-V4-Pro activates 49B of 1.6T parameters. Compute stays cheap, but every expert still has to be in memory, which is exactly why the weights column above is so brutal.

So the frontier models are architecturally impressive and practically unreachable. Both things are true, and the second one is what determines your shopping list.

The Tier That Actually Matters

Here’s the part worth bookmarking. These are Chinese open-weight models that run on hardware people actually own, at Q4 with 8K context:

Model	VRAM	Runs on
GLM-4 9B	6.9 GB	8GB card (Arc A750, RTX 3060 8GB, RTX 4060)
Qwen3 8B	7.0 GB	8GB card
Qwen3 14B	11.3 GB	12GB card (RTX 3060 12GB, Arc B580)
Kimi VL A3B (vision)	13.2 GB	16GB card
GLM-5 27B	20.8 GB	24GB card (RX 7900 XTX, RTX 4090)
Qwen3 30B-A3B MoE	21.8 GB	24GB card
Qwen3 32B	25.1 GB	32GB card, or 2× 16GB

Qwen3 8B at 7.0 GB is the headline. It fits an eight-gigabyte card, the kind that ships in mainstream gaming laptops, with room for a normal context window. That is the same open-weight ecosystem producing the 753B monster, and it’s the end of it most people will actually use.

The MoE entry is worth a second look too: Qwen3 30B-A3B needs 21.8 GB, a 30-billion-parameter model on a single 24GB card, because only 3B parameters activate per token while the memory cost stays bounded. That’s the sweet spot the architecture trend is opening up.

If your card isn’t on that list, check what fits yours directly. It scans every model against your actual pool and tells you which quantization level clears it.

Kimi K3 Landed, and It Broke the Pattern

Moonshot published Kimi K3’s open weights on July 27, 2026: roughly 2.8 trillion parameters with 104 billion active per token, a 1M-token context, and native MXFP4 weights occupying 1.561 TB on disk. It is the largest open-weight model released to date, and it sits far above everything in the frontier table above.

Our calculator puts full residency at 1,796 GB at Q4_K_M and 8K context. The smallest configuration in our GPU catalog that holds it is five Instinct MI430X cards, a 2,160 GB pool at 98% capacity. Six gives you room to actually work in.

The architecture is where it gets interesting, because K3 does not cache attention the way anything else on this page does. Of its 93 layers, only 24 use MLA, the compressed latent cache described earlier. The other 69 use Kimi Delta Attention, which stores no per-token keys and values at all. It advances a fixed recurrent state instead.

That distinction shows up directly in the numbers. A normal KV cache grows with every token you feed it. A recurrent state does not. K3’s cache is 0.7 GB at 8K context and 29.5 GB at 1M, so the total requirement moves from 1,796 GB to 1,829 GB. A 128-fold increase in context costs 1.7% more memory.

Hold that against a conventional 32B model, which spends 2.1 GB of cache at 8K alone. This is the clearest demonstration yet of where attention design is heading: caches that stop scaling with context, the same direction GLM-5.2’s shared indexers and DeepSeek V4’s compressed attention point in.

None of it helps the weights, though, and the weights are the wall.

K3 was enterprise hardware at every quantization we could model. That changed on July 29, when a 1-bit build brought it to 594 GB, though every published speed figure is still on datacenter hardware.

Our full breakdown of what Kimi K3 requires has the per-quantization figures and the GPU pools.

If You Can’t Fit It, You Have Options

Nothing above says a big model is impossible on small hardware, only that it’s impossible the straightforward way. Two escape hatches, both covered in depth:

CPU expert offloading. Because MoE models activate only a fraction of their experts per token, most of the weights can live in system RAM. A 118B model runs on a 12GB card this way, and a 26B multimodal model runs on an 8GB RTX 4060, with real trade-offs in speed, and real hidden costs in RAM and CPU that those articles spell out.
Rent before you buy. If you want to try a frontier model before committing to hardware, renting time on a big GPU costs far less than being wrong about a purchase.

Our calculator deliberately computes the full-residency number: everything on the GPU, maximum speed, no offloading. That’s the honest baseline, and it’s the number that tells you whether you need any of these techniques at all.

What to Watch

What K3 starts. Recurrent attention that does not grow with context is now shipping in a frontier model. Expect the next wave of releases to copy it, and expect advertised context lengths to stop tracking memory cost.
Licenses, not just scores. GLM-5.2 ships MIT, which is unusually permissive for a flagship. Others are more restrictive. If you’re building on a model commercially, the license matters more than a benchmark position.
Independent scores. Leaderboard claims move fast and are often vendor-supplied. Wait for third-party evaluations and the official model card before treating a ranking as fact.
Quantization is where the real gains are. Every number above is Q4. Dropping quality tiers or quantizing the KV cache changes what fits dramatically. Our flags guide covers the trade-offs.

DeepSeek shipped the official DeepSeek-V4-Flash-0731 on July 31, 2026, and it changes the bottom of this table more than the top. It needs 162 to 178 GB at 4-bit, roughly a tenth of what Kimi K3 asks for, which puts a model in this class within reach of a single well-specified machine for the first time.

FAQ

What are the best Chinese open-weight AI models right now?

The most prominent are GLM-5.2 (Z.ai, 753B, MIT licensed, 1M context), DeepSeek V4 (Pro at 1.6T total / 49B active, and Flash at 284B / 13B), the Qwen3 family from Alibaba (spanning 0.6B to 480B), and Moonshot’s Kimi K2 line (1T / 32B active), and Kimi K3 at ~2.8T with 104B active, released July 27, 2026.
Chinese labs currently occupy most top spots on open-weight rankings.

How much VRAM does GLM-5.2 need?

Roughly 526 GB at Q4 quantization with everything held in VRAM (about 457 GB of weights plus overhead), which needs a multi-GPU datacenter setup such as 4× B200. Its KV cache is remarkably small for its size (under 1 GB at 8K context) because it uses a compressed latent cache and shares attention indexers across layers.

Which Chinese open-weight model can I run on 8GB of VRAM?

Qwen3 8B (about 7.0 GB at Q4) and GLM-4 9B (about 6.9 GB) both fit an 8GB card with a normal context window. For 12GB, Qwen3 14B needs about 11.3 GB. For 24GB, Qwen3 30B-A3B MoE (21.8 GB) or GLM-5 27B (20.8 GB) fit comfortably.

How much VRAM does Kimi K3 need?

Kimi K3’s open weights were released on July 27, 2026, at roughly 2.8 trillion parameters with 104 billion active per token. Full VRAM residency runs to about 1,796 GB at Q4 quantization and 8K context, which is datacenter hardware. A 1-bit build released on July 29 brings it to 594 GB, reachable on a large multi-GPU machine, though no consumer-hardware speed figure has been published. Its 93 layers use Kimi Delta Attention on 69 of them and MLA on the other 24, and the KDA layers hold a fixed recurrent state rather than a per-token KV cache, so its cache is only 0.7 GB at 8K and 29.5 GB at a full 1M-token context.

Why do huge models have smaller KV caches than small ones?

Because KV cache size depends on attention architecture, not parameter count. Models like GLM-5.2, DeepSeek V4, and Kimi K2 use compressed latent caches, storing one small vector per layer instead of full per-head keys and values, plus sparse attention. A 753B model can therefore use under 1 GB of KV cache at 8K context while a 32B model with conventional attention uses 2.1 GB. The weights are what make the big models unreachable, not the cache.