The Chinese labs shipped a run of strong open-weight models between April and June 2026. They don't do the same thing well, so the useful question isn't which one is best overall. It's which one fits the job in front of you. Here's how they actually sort out, with the benchmarks and the cost.
The short answer
If you want one name, it's GLM-5.2. Z.ai released it on June 13, and it became the first open-weight model to lead the Artificial Analysis Intelligence Index: 51, fourth overall, behind only Fable 5, Opus 4.8 and GPT-5.5. It scores 62.1 on SWE-bench Pro, runs a million-token context, and ships under MIT. We cover it in full in the GLM-5.2 deep dive.
That's your default. The rest depends on what you're doing: DeepSeek V4-Pro still wins raw algorithmic coding, V4-Flash wins on cost, and Kimi wins on long-context tool use.
*Update, July 17, 2026:* the capability crown moved. Kimi K3 (released July 16) now ranks #3 on the Artificial Analysis index overall — above every other Chinese model and behind only Fable 5 and GPT-5.6 Sol. It is also priced like a flagship at $3/$15, and its open weights are committed but not yet released. Translation for this guide: K3 is the new ceiling; GLM-5.2 is still the value default. Both readings are below.
DeepSeek V4-Pro: the coding and reasoning flagship
A 1.6T-parameter model with 49B active. It tops LiveCodeBench at 93.5 and holds a 3206 Codeforces rating, which is grandmaster territory and ahead of several closed APIs. On agent tasks it scores 67.9% on Terminal-Bench 2.0, and on hard reasoning 90.1% on GPQA Diamond.
This is the model for serious code generation, competitive-programming-grade logic, and agent loops where quality decides the outcome. On Turiloop it costs a fraction of closed-model rates, currently with a discount on every model. See current pricing.
DeepSeek V4-Flash: the cost killer
284B total, 13B active — and since the stable 0731 build (shipped July 31) it no longer gives up much: a full retrain lifted it to 82.7 on Terminal-Bench 2.1 and near-Opus-4.8 territory on Agents' Last Exam, still at roughly $0.14 / $0.28 per million tokens with cache hits near $0.014.
Point it at the work that doesn't need deep reasoning, which covers most high-volume jobs: classification, extraction, summarization, first drafts. On Turiloop the per-token cost rounds to nothing.
GLM-5.2: the new overall leader
Z.ai's GLM-5.2, from June 13, is the strongest open-weight model on the market. It scores 51 on the Artificial Analysis Intelligence Index (first among open weights) and 62.1 on SWE-bench Pro, beating DeepSeek V4-Pro's 55.4 on multi-file GitHub issues. It carries a million-token context and MIT-licensed weights. Under the hood it's a 744B MoE with about 40B active, which is how you get frontier-adjacent quality without frontier pricing.
Make it your default for serious coding and long-horizon agents. It's now both the all-rounder and the model to beat. If you don't need that edge, the older GLM-5.1 is a cheaper everyday option. On Turiloop you pay as you go with an international card. Full breakdown in the GLM-5.2 deep dive.
Kimi K3: the new capability ceiling
Moonshot's K3, released July 16, is a 2.8T-parameter MoE with a 1M context that lands #3 on the Artificial Analysis index — the strongest Chinese model on aggregate, with the best published BrowseComp score (91.2) and a long-horizon agent Elo behind only Fable 5. It costs $3/$15 per million (cached input $0.30), roughly twice GLM-5.2 per input token and more than triple per output token, and it runs a single "max" reasoning level that burns real thinking tokens. Reach for it when the job is browse-heavy, tool-heavy, long-horizon agent work that the cheaper tier fumbles. Full specs, benchmarks and the honest cost math in the K3 launch breakdown.
Kimi K2.6: the long-context agent
Moonshot built K2.6 for long documents and multi-step agents. It scores 66.7% on Terminal-Bench 2.0 and about 58.6% on SWE-Bench Pro, and its prompt-cache pricing (around $0.16/M on cache reads) makes a long, reused system prompt cheap to carry across a session.
Reach for it on large-context analysis, agentic pipelines, and anything that reuses a big stable prompt. On Turiloop it's cheaper than any closed frontier model.
MiniMax M2.5: the value generalist
MiniMax's M2 series is a cost-efficient generalist for chat and assistant workloads, and a low-cost all-rounder on Turiloop. Use it for general assistant traffic when you want something cheap and capable and don't want to think hard about the choice.
One line per job
- Best value default, coding and agents → GLM-5.2
- Capability ceiling, hardest agent workflows → Kimi K3
- Hard algorithmic coding → DeepSeek V4-Pro
- High-volume cheap tasks → DeepSeek V4-Flash
- Everyday coding on a budget → GLM-5.1
- Long context and prompt reuse → Kimi K2.6
- General assistant traffic → MiniMax M2.5
The cheapest good setup is almost never one model. It's routing each job to the one that fits. On Turiloop they all sit behind one OpenAI-compatible key, so switching is a one-line model change, paid as you go with an international card.