Tech Stack
The most important model release of June wasn't from Anthropic or OpenAI. It was Zhipu's GLM-5.2 — a 753-billion-parameter open-weights model, MIT licensed, 1M-token context, that on FrontierSWE (the benchmark for hours-long autonomous coding tasks) trails Claude Opus 4.8 by a single percentage point.
Read that again: the gap between free-to-download and the $5/$25 frontier workhorse is now one point on long-horizon coding. The open-source tier list has been rewritten, so here's the state of it — including what I actually run on my own machines every day.
The new open-weights tier list (July 2026)
1. GLM-5.2 — the new open king. 753B parameters, 1M context, selectable reasoning modes, and a genuinely unrestricted MIT license. On the Intelligence Index v4.1 it scores 51 against MiniMax-M3's 44, DeepSeek V4 Pro's 44, and Kimi K2.6's 43. Zhipu positions it between Opus 4.7 and 4.8 on several evals, and VentureBeat's framing was blunt: beats GPT-5.5 on multiple long-horizon coding benchmarks at a sixth of the cost. The architectural trick is "IndexShare" — reusing the attention indexer across every four sparse attention layers — which is part of how a 753B model stays servable.
2. DeepSeek V4 Pro — the daily driver. The best all-rounder in open weights: tops open leaderboards on both agentic coding and graduate-level reasoning, MIT licensed, and the higher-effort V4-Pro-Max configuration hits 80.6% on SWE-Bench Verified. This one I don't have to speculate about — my personal agent stack runs a DeepSeek V4 Flash → Pro → Kimi escalation chain around the clock for briefings, email drafting, and routine automation. It's been the best cost-to-competence ratio in the market for months.
3. Kimi K2.7 Code — the coding specialist. Moonshot's fresh drop, posting +21.8% over K2.6 on Kimi Code Bench v2. K2-family models are the fallback tier in my own chain, and the trajectory between releases is steep.
4. Qwen3-Coder-480B — the enterprise-safe pick. 69.6% SWE-Bench Verified under Apache 2.0. If your legal team reads licenses (they should), Apache 2.0 and Gemma's terms are the friction-free paths into commercial products.
5. MiniMax M3 — the value option. 59.0% on SWE-Bench Pro, worth watching but a tier below the leaders.
When open weights actually win (and when they don't)
I ship on both closed APIs and open weights, so this is a practitioner's split, not ideology:
Open wins when: data can't leave your infrastructure (my on-device transcription product exists because of exactly this); volume is huge and margins matter — a sixth of frontier cost compounds fast at scale; you need to fine-tune on proprietary data; or you're in a market where API access is unreliable.
Closed still wins when: you need the absolute frontier (Fable 5 and Sol remain unmatched at the very top); your team is small and ops time is your scarcest resource — self-hosting a 753B model is a real engineering project, MIT license or not; or your spend is too small for infrastructure to pay back. Under roughly $2-3K/month in API spend, self-hosting big models is usually a hobby, not a saving.
The hybrid pattern that actually works
The setup I run and recommend: closed frontier models for engineering and architecture (Claude Code on Opus/Fable), open-weights chains for high-volume routine work (DeepSeek → Kimi via API providers at open-model prices), and truly local models only where privacy is the product. You don't have to host GLM-5.2 yourself to benefit from it either — every major inference provider is racing to serve it, at prices that undercut closed workhorses badly.
What GLM-5.2 changes strategically
Every previous "open source is catching up" moment came with an asterisk — okay on benchmarks, brittle on long tasks. A one-point gap on hours-long autonomous coding removes the asterisk. Closed labs now have roughly a six-month capability lead to justify a 6x price gap, and that math gets harder to defend every quarter. For builders, the winning move hasn't changed, it's just gotten more profitable: route by tier, benchmark on your own workload, and let the price war — closed vs closed, and now open vs closed — pay you.
Next on my list: GLM-5.2 as the workhorse slot in my agent chain. If it holds up in production the way it benchmarks, that's another 60% off the token bill. I'll publish what I find.
Want to build something like this?
I architect and deploy end-to-end AI systems — from MVP to revenue.
Let's TalkOr ask Angelina — my AI twin in the bottom-right corner. She knows my full build history, live GitHub, and how I'd approach your project.