Aug 29, 2026
Choosing an Open-Weight Model in 2026: DeepSeek vs Qwen vs Kimi vs GLM vs Mistral
Six open-weight releases landed in one August week, and GLM-5.3's full weights plus Tencent Hy4 preview followed this week. A decision guide by workload: long context, tool use, multimodal, reasoning, and self-hosting.
Six open-weight releases landed in the same week of early August. This week, Zhipu put GLM-5.3's full weights on Hugging Face and Tencent open-sourced Hy4 preview. The field is wide. The hard part is no longer finding an open model — it is choosing one without renting a datacenter or believing a press release.
What shipped, as verified
On August 3, 2026, five labs shipped open-weight updates on one day — DeepSeek V3.2, Qwen3-Omni, Mistral Magistral, Kimi K2, and GLM-5.2 — with Hugging Face's open-r1 project as the sixth release, completing the cluster. On August 27-28, GLM-5.3's weights landed on Hugging Face and Tencent released and open-sourced Hy4 preview.
The table sticks to what each lab's own release materials state. Where a claim is vendor-reported and self-run, the Standout column says so.
| Model | Lab | Stated in the release notes | Standout |
|---|---|---|---|
| DeepSeek V3.2 | DeepSeek | Sparse MoE routing refinements, 128K context with sliding-window extension | Routing efficiency over raw size |
| Qwen3-Omni | Alibaba | Vision + audio + text under an open-weight license, with SFT/DPO recipe notes | Multimodal input in one open model |
| Magistral | Mistral | Open-weight reasoning model with a chain-of-thought template and Le Chat integration | Reasoning with visible steps |
| Kimi K2 | Moonshot | 256K-token context window and tool-use support | Long context plus tool calls |
| GLM-5.2 | Zhipu | Compressed-attention MoE (78 layers, hybrid local + compressed global attention) | Attention architecture |
| open-r1 | Hugging Face | Full open weights, training data, and training code for a frontier-class reasoning reproduction | Reproducibility |
| GLM-5.3 | Zhipu | Full weights + BF16 on Hugging Face (~753B per repo metadata, 2026-08-28); Flash variant ~321B (2026-08-27) | Same base as GLM-5.2; all gains from post-training (vendor-stated) |
| Hy4 preview | Tencent | 770B total / 49B active MoE, context beyond 1M tokens (2026-08-28) | Coding, office productivity, research focus |
Decision paths
"I need long context and tool calls" — Kimi K2
Moonshot's release notes give Kimi K2 a 256K-token context window with tool-use support. That combination fits pipeline work: digest building, cross-document summarization, agent loops that read more than they write. DeepSeek V3.2's 128K context with sliding-window extension covers most single-project corpora, with less headroom.
"I need multimodal input" — Qwen3-Omni
Qwen3-Omni is the only model in this set whose release notes document vision, audio, and text together under an open-weight license. If the workflow ingests screenshots, podcast audio, or slide decks, start there before bolting a separate speech or vision model onto the stack.
"I need reasoning I can inspect" — Magistral, or GLM-5.3 with caveats
Magistral ships a chain-of-thought template, so the reasoning surface is part of the product. GLM-5.3 is the stronger coding claim — Z.ai's launch post claims a 50% improvement over GLM-5.2 on its in-house Z.ai Code Bench — but that benchmark is self-run, and no independent replication is cited. If you pick GLM-5.3, quote the claim as Z.ai's own.
"I want the lowest verified price" — check the price page, then Hy4 preview's numbers
The only price in this field quotable from a primary source today is Tencent's preview pricing for Hy4: USD 0.834 per million input tokens, USD 2.501 per million output tokens, and USD 0.042 per million cache-hit tokens. It is preview pricing, alongside a two-week free window in WorkBuddy and CodeBuddy. Every other vendor's price in this guide is deliberately unstated — open the pricing page the day you choose, and quote what it says then.
"I want to run the weights myself" — match the hardware budget first
GLM-5.3's full weights are around 753B parameters. That is datacenter-class; the Hugging Face index lists hosted providers for it (together, baseten, zai-org for the full model; novita, fireworks-ai, baseten for Flash). For genuine self-hosting, GLM-5.2-class models and the open-r1 stack — which publishes training data and code, not just weights — are the honest starting points. License terms vary by model and were not verified for this guide; check the model card before you promise anything to an audience.
Common mistakes
- Treating vendor benchmarks as independent. Every capability number in the August 3 cluster claims is vendor-reported, and GLM-5.3's coding and cyber benchmarks are Z.ai's own runs.
- Reading parameter counts as runnability. "Open weights" does not mean "runs on my machine" at 753B; check the parameter count and the hosted-provider list before demoing a local setup.
- Reading context windows as quality. 256K or 1M tokens describe capacity, not output quality — capacity only matters when the workload actually fills the window.
- Quoting prices from memory. Preview pricing is promotional and time-boxed; a price quoted from an old video is a correction waiting to happen.
- Picking by total parameter count. Hy4 preview's 49B active parameters (out of 770B) change the serving math more than the total does. Active-parameter counts are the number that maps to cost.
For the layer above the model, see Choosing an Agent Framework in 2026. For this week's releases in detail, see Tencent Hy4 preview and GLM-5.3 open weights.