Kimi K3 vs DeepSeek V4 vs GLM-5.2: Which Chinese Open Model Wins?

Short answer

• Kimi K3: the capability pick. Highest independent score (Artificial Analysis Intelligence Index 57), 1M context in and up to 1M tokens out, native multimodal. Also the most expensive, and impossible to self-host at home.
• GLM-5.2: the price pick. MIT license, $1.40/$4.40 per million tokens as of August 2026, 1M context, and the only trillion-scale model here that a 256GB Mac Studio can technically load.
• DeepSeek V4-Flash: the operability pick. 284B total / 13B active, MIT license, 155GB at 4-bit — the only one of the three that runs well on hardware a person owns. Full breakdown in DeepSeek V4-Flash vs gpt-oss-120b.

Ready to buy? See the tested hardware list with current prices.

Specs that matter

Capability: K3 leads, but reads the fine print

On the independent Artificial Analysis Intelligence Index, K3 scores 57, GLM-5.2 Max scores 51, and V4-Flash-0731 scores 50. GLM-5.2 topped the open leaderboard at its June launch; K3 has since passed it.

On coding agents, aggregated vendor-reported numbers show K3 at 67.5 on DeepSWE versus 46.2 for GLM, and 88.3 versus 82.7 on Terminal-Bench 2.1. Label those what they are: vendor launch numbers, collected by a secondary aggregator.

Architecture explains some of the gap. K3 introduces Kimi Delta Attention, Attention Residuals, and Stable LatentMoE — Moonshot claims roughly 2.5x scaling efficiency over K2 — and it takes text, image, and video input. GLM-5.2 is text only. K3 is also quantization-aware trained: MXFP4 weights with MXFP8 activations from SFT onward.

Price: GLM undercuts hard, but watch token burn

As of August 2026, GLM-5.2 lists $1.40 per million input tokens and $4.40 output (cached input $0.26). K3 lists $3 and $15 (cached input $0.30).

Concrete example: a 1M-input, 100K-output agent job costs about $1.84 on GLM-5.2 versus $4.50 on K3.

But sticker price is not spend. Hacker News users report K3 ends up more expensive per task than Opus 4.8/5 and GPT 5.6 Sol “because it chews so much” — it burns reasoning tokens. One commenter counters that GLM-5.2 is even worse on token efficiency. Measure your own workload before you commit either way.

GLM’s price has an architectural basis: its IndexShare sparse attention reuses indexers across layers and cuts per-token operations 2.9x at max context, per the official model card.

Self-hosting: only one of these fits in a house

Kimi K3: no. MXFP4 weights are about 1.4TB (FP16 is ~5.6TB). The practical minimum is an 8-node cluster of 8x 80GB GPUs — 64 GPUs. This is a datacenter model with open weights, not a local model.

GLM-5.2: barely. Unsloth’s 2-bit dynamic quant compresses the 1.51TB original to 239-241GB. A 256GB M3 Ultra Mac Studio runs it at roughly 3-9 tok/s. Q4 is ~370-475GB depending on the build. It works as a demo, not as a daily driver.

DeepSeek V4-Flash: yes. The UD-Q4_K_XL GGUF is 155GB (~162GB in memory) — comfortable on a 192GB machine. A 3-bit build (~103GB) runs on 128GB boxes like the ones in our 128GB RAM guide. llama.cpp has supported it in mainline since July 2026, with DSpark speculative decoding merged August 2.

If self-hosting matters to you, the decision collapses to V4-Flash before capability even enters the picture.

Decision table

Final recommendation

Use GLM-5.2 as the default API workhorse: MIT, cheap, 1M context. Escalate hard tasks to Kimi K3 and accept the token bill. Run DeepSeek V4-Flash if the weights must live on your own hardware — it is the only flagship here that meets a home machine halfway.

Next steps

• DeepSeek V4-Flash vs gpt-oss-120b
• Best local LLMs for 128GB RAM
• Best local models for OpenClaw

Sources

• Kimi K3 model card (Hugging Face)
• GLM-5.2 model card (Hugging Face)
• Artificial Analysis Intelligence Index
• Unsloth dynamic quants documentation

Before you order parts, check the tested hardware list for current prices by tier.