GLM-5.2: 744B Open Weights Top the Intelligence Index

What builders need to know

• Open weights took the lead. Released on 13 June 2026, GLM-5.2 from Z.ai is now the top open-weight model on the Artificial Analysis Intelligence Index (v4.1), scoring 51 — clear of DeepSeek V4 Pro at 44 and Kimi K2.6 at 43.
• It is a sparse giant. 744B total parameters but only around 40B active per token, thanks to a Mixture-of-Experts design, with a 1M-token context window.
• The licence is the headline. The weights ship under the permissive MIT licence — self-hostable, fine-tunable and commercially usable, with no per-token API bill attached.
• Coding at a fraction of the cost. Z.ai reports GLM-5.2 beating GPT-5.5 on several coding benchmarks at roughly one-sixth of the cost — a vendor claim worth validating, but a striking one.
• Long context got cheaper to serve. A new "IndexShare" sparse-attention mechanism cuts per-token compute by about 2.9x at a 1M-token context, which is where self-hosting economics usually fall apart.

For the past two years the open-weight camp has been the pragmatic choice rather than the best one: you accepted a capability gap in exchange for control, portability and a bill you could actually predict. GLM-5.2 narrows that trade to the point where it is worth re-examining. When an independent benchmark like the Artificial Analysis Intelligence Index puts a downloadable, MIT-licensed model at the front of the open-weight field, the question for a builder in Bengaluru or Bristol stops being "is open good enough?" and becomes "at what volume does running this myself beat renting it?" That is a spreadsheet question, not a faith question, and it has a clean answer.

What GLM-5.2 actually is

Strip away the leaderboard drama and the architecture is the interesting part. GLM-5.2 is a Mixture-of-Experts (MoE) model: 744B parameters in total, but a router selects only a small subset of "experts" for each token, so roughly 40B parameters are active on any given forward pass. That is the same design philosophy behind other recent open-weight heavyweights — we covered the shape of it when NVIDIA's Nemotron 3 Ultra shipped as a 550B open-weight MoE — and the practical consequence is the same: you get the knowledge capacity of a very large model with the inference throughput of a much smaller one. Latency and cost track the ~40B active count, not the 744B headline.

The second notable piece is context. GLM-5.2 carries a 1M-token window, which on its own is table stakes in mid-2026, but Z.ai paired it with a new sparse-attention scheme it calls IndexShare that reportedly cuts per-token compute by around 2.9x at full 1M-token context. Long context is where self-hosting maths usually collapses, because attention cost grows quickly with sequence length and the GPU-hours pile up. Cutting that compute by nearly three times is not a marketing footnote; it is the difference between a 1M-token feature you can afford to run and one you switch off in production. The efficiency work is also a reminder that the open-weight frontier is not just chasing scores — the same instinct shows up in research like on-policy distillation pushing frontier reasoning into small models, where the win is capability per unit of compute rather than raw size.

The third piece — and the one that makes all of this actionable — is the licence. Z.ai has released the GLM-5.2 weights under the MIT licence. That is about as permissive as open licensing gets: you can download the weights, host them wherever you like, fine-tune them on your own data, ship them inside a commercial product and charge for it, with no royalty and no usage clause waiting to bite. Compared with the more restrictive community licences that some open-weight releases carry, MIT removes the legal ambiguity that makes procurement and legal teams nervous.

Open weights, not open source — the distinction that matters

It is worth being precise here, because the two phrases get used interchangeably and they are not the same thing. GLM-5.2 is open weights, not fully open source. Z.ai has released the trained model — the weights you load onto a GPU — under MIT. What it has not released is the complete training-data pipeline, the data itself and the full recipe needed to reproduce the model from scratch. In the strict sense used by the open-source software world, a model is only "open source" when you could rebuild it end to end from published components.

Does that distinction cost you anything as a builder? For the vast majority of use cases, no. You are not going to retrain a 744B model from raw data — almost nobody outside a frontier lab is. What you need is the legal right to run, adapt and deploy the weights, and MIT gives you exactly that. The place the distinction bites is narrower: you cannot fully audit what went into the training set, so provenance and data-governance questions cannot be answered by inspecting a public corpus. For a UK team under strict procurement rules, or an Indian enterprise with a data-lineage requirement, that is a real limitation to log — not a reason to avoid the model, but a caveat to document. Being accurate about "open weights" versus "open source" is also just good hygiene; overstating it in a client deck is the kind of thing that gets found out.

Reading the leaderboard: GLM-5.2 vs the open-weight field

The Intelligence Index result is best understood in context rather than as a single number. Here is where GLM-5.2 sits against the two open-weight models it displaced, on the dimensions a self-hosting builder actually cares about — capability, licence terms, context length and the rough cost posture of running each one.

Intelligence Index scores per Artificial Analysis (v4.1). Licence, context and cost-posture notes are indicative; verify against each model card before committing to a deployment.

Two things stand out. First, the seven-point gap between GLM-5.2 and DeepSeek V4 Pro is meaningful on this index — it is not a rounding difference between near-identical models. Second, GLM-5.2's MIT licence is arguably as important as its score: a permissive licence removes friction that a restrictive community licence adds at exactly the moment you want to ship. If you are choosing a coding-first open model, the trade is different again — Kimi's lineage is deliberately agentic, and the K2.7 Code release built for MCP and agentic coding may still be the better fit for a tool-calling harness even at a lower index score. The leaderboard tells you general capability; your workload tells you which model to actually run.

The self-hosting maths for Indian and UK builders

Here is the decision that actually matters. A 744B model sounds unaffordable to run, and if it were dense it would be. But because only ~40B parameters are active per token, throughput behaves like a far smaller model — the cost is memory, not compute. The full weights must be resident in GPU memory, which in practice means a multi-GPU node: an 8x H100 or H200 host at quantised precision is the usual starting point for serving GLM-5.2 yourself. That is a serious box, whether you rent it from AWS Mumbai, Azure UK South, an Indian neocloud, or a UK sovereign-compute provider.

The question is not "can I afford one node?" but "will I keep it busy?" An idle 8-GPU host is the most expensive thing in your budget. The rough shape of the trade looks like this:

# Self-host vs API — the utilisation break-even (illustrative)
#
# Fixed cost of an 8x H100/H200 node (rented, per month):  substantial and CONSTANT
#   -> you pay it whether the GPUs run at 5% or 95% utilisation
#
# API cost: pay-per-token, scales with usage, near-zero when idle
#
# Break-even heuristic (validate with YOUR numbers):
#   sustained GPU utilisation  < ~30-40%   -> hosted API is usually cheaper
#   sustained GPU utilisation  > ~30-40%   -> self-hosting a 744B MoE starts to win
#
# Levers that move the line in your favour:
#   - IndexShare -> ~2.9x less compute at 1M ctx (long-context work gets cheaper)
#   - quantisation -> fewer GPUs to hold the weights
#   - batching -> higher utilisation from the same hardware
#   - in-region hosting -> avoids cross-border data-residency cost/risk

For an Indian startup running steady, high-volume inference — document processing, support automation, code assistance across a large team — a well-utilised self-hosted GLM-5.2 can undercut a per-token API decisively, and it keeps regulated data inside an Indian region for DPDP purposes. For a UK team handling personal or regulated data, self-hosting on a UK South region or a sovereign-compute route is often the cleaner compliance story, with the added benefit of escaping US-API lock-in entirely. But for a team with spiky, low or unpredictable traffic, the same node sits idle most of the day and a hosted endpoint is simply cheaper. The model being free to download does not make the GPUs free to rent.

Where GLM-5.2 does not change your plan

A leading open-weight model is genuinely useful, but it does not rewrite every decision. If your product needs a small model that runs on a laptop or a phone, a 744B MoE is the wrong tool no matter how well it scores — that is the territory of quantised on-device models like Gemma 4's QAT builds running in about 1GB on Arm laptops, and GLM-5.2 does not compete there. If your traffic is genuinely low, the operational overhead of running and updating a multi-GPU inference stack — monitoring, autoscaling, security patching, model updates — is real work that an API abstracts away for you. And if your workload is dominated by a single narrow task, a smaller fine-tuned model may beat a general 744B giant on both cost and quality.

The honest framing is optionality. GLM-5.2 does not oblige you to self-host; it makes self-hosting a credible option for the first time at the top of the open-weight field, under a licence permissive enough that your legal team will not block it. That is a meaningful shift in negotiating position for builders who have spent two years choosing between expensive closed APIs and a capability gap. You can browse the wider open-weight landscape on our open-source news stream, but the pattern is consistent: the gap between open and closed is closing from the open side, and the leverage is moving towards the teams that can run their own stack.

The bottom line

GLM-5.2 is the clearest evidence yet that the leading edge of open weights is no longer a compromise. An independent benchmark ranks it first among open-weight models; a permissive MIT licence removes the legal friction; a sparse MoE architecture keeps inference affordable; and IndexShare makes even the 1M-token context economical to serve. For Indian and UK builders the practical move is not to switch everything to GLM-5.2 tomorrow — it is to run the utilisation maths honestly, keep your inference layer portable, and be ready to move steady high-volume, data-sensitive workloads onto a self-hosted tier the moment the numbers cross the line. The model is free to download; the discipline is in knowing when the GPUs are worth it. Full benchmark detail is in the Artificial Analysis writeup, with additional context in Implicator's coverage.