For most of the AI era, “open” meant a capable-but-second-tier model you ran because you could not afford the frontier. That changed in 2026. A cluster of Chinese labs — Moonshot AI, Alibaba, DeepSeek, Zhipu (Z.ai), ByteDance and Tencent — now ship open-weight models that trade blows with the best closed systems on coding, reasoning and agentic benchmarks, and they release the weights under permissive licences you can actually deploy. On several 2026 leaderboards, four of the top five open models come from Chinese labs. This guide is the map: who ships what, and what it means if you run a business in Malaysia.
The 2026 landscape at a glance
| Lab | Flagship (2026) | Architecture | Context | Licence | Best for |
|---|---|---|---|---|---|
| Moonshot | Kimi K3 | ~2.8T MoE | 1M | Modified MIT (open weights) | Autonomous coding agents |
| Alibaba | Qwen 3.6 | 27B dense / 35B-A3B MoE (Max closed) | 256K→1M | Apache-2.0 | Multilingual + easy self-host |
| DeepSeek | V4 (pro / flash) | MoE, adaptive reasoning | 1M | Open weights (MIT-style) | Reasoning + cost efficiency |
| Zhipu / Z.ai | GLM-5.2 | ~753B MoE / ~40B active | 1M | MIT | Long-horizon coding agents |
| ByteDance | Doubao-Seed 2.0 / Seed 2.1 | Mostly served via API | Large | Mostly proprietary; partial open (Seed-OSS) | Cheapest API, multimodal |
| Tencent | Hunyuan Hy3 | 295B MoE / 21B active | 256K | Apache-2.0 | Multimodal (image/3D/video) |
Why “open weights” is the story of 2026
An open-weight model is one whose trained parameters you can download and run on your own hardware or a cloud of your choosing. That is different from a closed API model (GPT-5, Claude, Gemini) where the weights never leave the vendor. It is also subtly different from fully “open-source” — most of these labs release the weights and a licence, but not the full training data. For a buyer, the practical wins are the same: no per-seat lock-in, the option to self-host sensitive data, predictable cost, and the freedom to fine-tune. Most of the 2026 flagships ship under MIT or Apache-2.0, which permit commercial use.
Moonshot — Kimi K2.6 and K3
Moonshot AI's Kimi line is the agentic-coding specialist. Kimi K2.6 is a roughly trillion-parameter mixture-of-experts model that ranked among the top open models on coding arenas; in July 2026 Moonshot released Kimi K3, a reported ~2.8-trillion-parameter model with a 1M-token context, billed as the largest open-weight model yet and tuned for long-running autonomous coding agents. Full breakdown of Kimi →
Alibaba — Qwen 3.6
Qwen is the most widely deployed open model family on earth, largely because it is small, permissively licensed and easy to run. The 2026 Qwen 3.6 series includes a dense 27B and a 35B-A3B MoE (only ~3B parameters active per token) that score at the frontier on agentic coding while fitting on modest hardware. Alibaba keeps its very top “Max” model closed, but open-sources the rest under Apache-2.0. Full breakdown of Qwen →
DeepSeek — V4
DeepSeek shocked the market in 2025 with cheap, strong reasoning models. In April 2026 it shipped DeepSeek V4, which merged its general (V) and reasoning (R) lineages into a single model that decides how hard to think per task — which is why there is no separate “R2” or “R4” product. V4 is a large MoE with a 1M context and famously low token cost. Full breakdown of DeepSeek →
Zhipu / Z.ai — GLM-5.2
GLM-5.2 (June 2026) is a ~753B-parameter MoE (~40B active) under the MIT licence, with a 1M-token context and a clear focus on long-horizon coding agents at roughly a sixth of the cost of closed rivals. Notably, the GLM-5 generation was trained largely on domestic Huawei Ascend silicon. Full breakdown of GLM →
ByteDance — Doubao and Seed
ByteDance is the outlier: its Doubao family is the cheapest and one of the most widely used in China, but it is served mainly through ByteDance's API rather than as open weights (the lab open-sources some smaller “Seed-OSS” models). Seed 2.1 (June 2026) is pitched as a reliable general agent. If you want the lowest API price and strong multimodal, Doubao is the pick; if you need downloadable weights, look elsewhere. Full breakdown of Doubao/Seed →
Tencent — Hunyuan Hy3
Tencent's Hunyuan Hy3 (HunYuan 3.0, July 2026) is a 295B MoE that activates just 21B parameters per token, released under Apache-2.0. Tencent positions it as matching agentic and coding performance of models several times its size at far lower cost per token — and Hunyuan is unusually strong on multimodal generation (image, 3D and video). Full breakdown of Hunyuan →
How a Malaysian business should choose
Four questions settle most decisions:
- Do you need to self-host? If you handle PDPA-sensitive data (finance, healthcare, government), open weights let you run the model inside your own VPC or on-prem, so customer data never leaves your control. Qwen (small, easy) and GLM (strong, MIT) are the friendliest starting points; Kimi K3 and DeepSeek V4 are heavier and need serious GPUs.
- Is it a coding or agent workload? Kimi K3, GLM-5.2 and DeepSeek V4 lead on autonomous, multi-step coding agents. Pair them with n8n workflows or Claude-style orchestration.
- How important is Bahasa Malaysia and Chinese? Chinese labs train heavily on Chinese and increasingly on multilingual data; Qwen and GLM are particularly strong across Chinese and Southeast Asian languages, which matters if your customers are not English-first. (See why language affects cost in our tokens guide.)
- API cost vs control? If you just want the cheapest capable API and do not need the weights, Doubao and DeepSeek are hard to beat on price. If you want portability and no lock-in, choose an MIT/Apache model you can host anywhere.
The caveats worth stating plainly
Benchmarks are marketing until you test on your task — always run a small internal eval before committing. Licences vary (some “open” releases carry use restrictions, so read them). And there are governance and geopolitics questions: some buyers, especially in government or regulated finance, will have policies about model provenance and where inference runs — self-hosting an open-weight model on infrastructure you control is often the cleanest answer. If you are building on these models, our AI Engineering course covers evaluation, self-hosting and putting open models into production. For the underlying concepts — parameters, context, MoE, fine-tuning — start with our LLM fundamentals and parameters & scaling laws explainers.
The benchmark scoreboard
Numbers cut through the marketing. Here are each flagship's headline coding scores — with one important caveat built into the caption below.
Headline coding benchmarks (2026)
| Benchmark | Score | What it measures |
|---|---|---|
| Kimi K3 | 76.8% | SWE-bench Verified |
| Qwen3.6-27B | 77.2% | SWE-bench Verified |
| DeepSeek V4 (Pro-Max) | 80.6% | SWE-bench Verified |
| Doubao-Seed 2.0 Pro | 76.5% | SWE-bench Verified |
| Hunyuan Hy3 | 74.4% | SWE-bench Verified |
| GLM-5.2 | 62.1 | SWE-bench Pro (harder test) |
Official resources & downloads
Go straight to the source — official sites, model cards and weight downloads: