Open-weight models are no longer the cheap alternative in 2026. They have become the standard. According to the Lushbinary website, which published a comparative guide in June, today there are five or six genuinely capable open models, all with MIT or Apache 2.0 licenses, for a fraction of the cost of closed models. The problem is no longer ‘to use or not to use’, but ‘which one to use for each job’.
The guide compares the current leaders: GLM 5.2 (Z.ai/Zhipu), DeepSeek V4 Pro and Flash, MiniMax M3, Kimi K2.6/K2.7 Code and Qwen 3.6 (Alibaba). Data comes from official technical sheets and provider pricing pages.
GLM 5.2: the new leader in raw intelligence
Released on June 13, 2026, GLM 5.2 is a 744 billion parameter MoE model, with a context of 1 million tokens and MIT license. It leads the Artificial Analysis Intelligence Index among open models, with a score of 51, ahead of MiniMax M3 (44) and DeepSeek V4 Pro (44). On the GDPval-AA v2 benchmark, it scores 1524 points. It surpasses GPT-5.5 on several long-horizon coding benchmarks, costing about one-sixth the price: $5.80 per million tokens vs $35 for GPT-5.5. It is one point behind Claude Opus 4.8 on FrontierSWE (74.4 vs 75.1) and MCP-Atlas (76.8 vs 77.8), and wins on AIME 2026, IMOAnswerBench and Terminal-Bench 2.1.
DeepSeek V4: the price floor
DeepSeek released V4 Pro and V4 Flash on April 24, 2026, both with MIT license and native context of 1 million tokens (up to 384K output). V4 Pro has 1.6 trillion parameters (49B active); Flash has 284B (13B active). V4 leads algorithmic reasoning benchmarks among open models: LiveCodeBench 93.5% (first globally), Codeforces 3206, HMMT 95.2%, GPQA 90.1%. Flash costs $0.14/0.28 per million tokens and achieves 79% on SWE-bench Verified.
MiniMax M3: long context and native multimodality
MiniMax M3 (June 1, 2026) offers 1 million token context, native multimodality (image + text) and modified MIT license. It uses the MiniMax Sparse Attention (MSA) architecture, which accelerates decoding by 15.6x and prefill by 9.7x in 1M context. Standard price: $0.60/2.40 per million tokens (cached input: $0.12). Ideal for document understanding, multimodal agents and long-context workflows.
Kimi K2.6 and K2.7 Code: stable agents
Moonshot developed Kimi K2.6 (~1 trillion parameters, 256K context, Apache 2.0) for long-horizon agentic tasks and multi-step tool use. K2.7 Code (June 13) is a coding-specialized variant that reduces thinking tokens by about 30%, cheapening long runs. Where GLM 5.2 wins on raw intelligence and DeepSeek on price, Kimi wins on agentic stability.
Qwen 3.6: the versatile for one GPU
Qwen 3.6 (Alibaba) is a compact MoE of 35B with 3B active, 256K context, Apache 2.0 license and vision support. Runs on a single GPU, with strong tool calling capability. Ideal for self-hosting, local workloads and hardware-constrained scenarios.
Summary comparison table
| Model | Params (active) | Context | License | API $/1M (in/out) | Multimodal |
|---|---|---|---|---|---|
| GLM 5.2 | 744B MoE | 1M | MIT | 1.40 / 4.40 | Text |
| DeepSeek V4 Pro | 1.6T (49B) | 1M | MIT | 1.74 / 3.48 | Text |
| DeepSeek V4 Flash | 284B (13B) | 1M | MIT | 0.14 / 0.28 | Text |
| MiniMax M3 | MoE (MSA) | 1M | MIT mod. | 0.60 / 2.40 | Yes (native) |
| Kimi K2.6 | ~1T MoE | 256K | Apache 2.0 | ~0.60 mixed | Text |
| Qwen 3.6 (35B-A3B) | 35B (3B) | 256K | Apache 2.0 | Local / low | Yes (vision) |
Decision framework
According to the Lushbinary guide, the choice depends on priority:
- GLM 5.2: best general intelligence and long coding; replaces GPT-5.5 or Claude with 6x savings.
- DeepSeek V4 Flash: minimum cost for routing, classification and simple coding.
- DeepSeek V4 Pro: maximum algorithmic reasoning with long context at the lowest price in its category.
- MiniMax M3: multimodality + long context with good cost-benefit.
- Kimi K2.7 Code: stable and cost-effective coding agents in thinking tokens.
- Qwen 3.6: self-hosting on single GPU, privacy and local workloads.
The open-weight ecosystem is mature. The frontier is no longer a closed club. It is up to the developer to choose the right tool for each task.