GLM-5.2 vs DeepSeek V4: Open-Weights Coding Kings

Two names dominate every “best open-weight coder” thread in mid-2026: GLM-5.2 from Z.ai and DeepSeek V4 Flash. Both publish downloadable weights, both serve at FP8, and both are priced to embarrass the closed frontier. They are also opposite bets — one a 744B-parameter Mixture-of-Experts flagship, the other a throughput machine priced like a utility. The arena runs them on identical prompts, and you can watch the duel unfold vote by vote: GLM-5.2 vs DeepSeek V4.

Open-model comparisons usually die on quantization. One side gets quoted from the vendor’s BF16 benchmark rig, the other from a community 4-bit quant, and the “comparison” ends up measuring the quantization lottery instead of the models. Not here: both models run at FP8 through the same routing, so a vote for one output over the other is a vote about the model.

One weight class, two business models

GLM-5.2 is Z.ai’s open-weights flagship: 744B parameters arranged as a Mixture-of-Experts, so only a slice of the network fires per token, with official API pricing at $1.40 in / $4.40 out per million tokens and weights you are allowed to take home. DeepSeek V4 Flash attacks from below: open weights served at commodity rates around $0.14 in / $0.28 out per million tokens, engineered so that calling the API is cheaper than thinking about alternatives.

• GLM-5.2 — frontier-scale MoE that happens to be open; priced to undercut closed flagships.
• DeepSeek V4 Flash — throughput-first economics; priced to compete with your electricity bill.
• Open weights on both sides, so neither locks you in if pricing, policy, or geopolitics shift.

Spec sheet

The stable facts are below. The performance question is deliberately absent, because that part is decided live by voters, not by a launch post.

What the arena actually measures

Every matchup here is community-voted and one-shot: both models get the same coding prompt, produce one attempt each, and real people pick the better result without knowing which model wrote what. No retry loops, no cherry-picked demos, no vendor-run harness. The results accumulate live, which is why this article quotes no scores of its own — the honest numbers live on the leaderboard, and the fastest way to form an opinion is to open the arena and cast a few votes yourself.

The self-host math, 744B edition

This is where the romance of open weights meets arithmetic. At FP8, 744B parameters is roughly 744 GB of weights before you allocate a byte of KV cache — multi-GPU node territory, not a workstation project. The MoE design keeps per-token compute manageable once the model is loaded, but loaded is the expensive word. For most teams, GLM-5.2’s open weights are less “run it tonight” and more an insurance policy: on-prem is realistic for enterprises with genuine clusters, and the mere existence of the weights disciplines every host’s pricing.

• Strict data residency plus a real GPU cluster: self-hosting GLM-5.2 is viable, and the open weights make it legal.
• Everyone else: both models are cheapest as hosted APIs, where V4 Flash’s utility pricing is brutal to compete with.
• Either way the exit door stays open — switch hosts, change quants, or go on-prem later without rewriting your stack.

With open weights, the question is never whether you can leave your provider — only whether it is worth the electricity.

Calling it

Do not call this one from a spec sheet. A 744B MoE and a flash-priced throughput model can trade wins prompt by prompt — dense logic one round, fast clean scaffolding the next — and the only honest record of who is ahead is the running vote. Watch the live head-to-head, weigh the standings against your own deployment reality, and if you are still mapping the wider field beyond these two, the best AI for coding guide covers how the open-weight leaders stack up against the closed frontier. For this particular crown, the community is still counting, one blind vote at a time.