Open-Weights Models of 2026
Compare current Chinese open-weight models by the artifact you can deploy: weight bytes, attention state, supported execution paths, and evaluation conditions. A dated guide to GLM, Kimi, DeepSeek, Qwen, and MiniMax, with worked sizing checks that separate parameter counts from GPU capacity.
TL;DR: A model name does not specify a deployment. Pin the checkpoint, inspect its stored tensors and attention implementation, then test a supported engine on your workload. The September 7, 2026 landscape includes trillion-parameter mixtures of experts and smaller dense models. Active parameters approximate arithmetic per token; they do not determine resident weight memory or guarantee throughput. Open weights also carry different licenses.
Start with the artifact you would actually load
A team asks whether its existing GPU node can serve a new release. In the table, "18B active" refers to a mixture-of-experts model: each token runs through only some of its expert blocks, so 18B parameters compute per token while all of them must still be stored. Before answering, distinguish the model's language-parameter count from all tensors in its checkpoint, including vision components, scales, and speculative modules. The table uses decimal GB and TB. Tensor bytes come from each repository's model.safetensors.index.json at pinned revisions linked in the size column, checked September 7, 2026. They exclude file headers and runtime allocations.
| Official checkpoint | Model scale | Stored tensors | Deployment distinction |
|---|---|---|---|
| GLM-5.3 | About 753B checkpoint parameters | 755.6 GB, mixed FP8/BF16 | Latent attention with sparse indexing; model-specific license |
| GLM-5.3-Flash | 320B, 18B active | 328.3 GB, mixed FP8/BF16 | Multimodal; linear and sparse attention; MIT |
| Kimi-K3 | 2.8T, 104B active | 1.561 TB, MXFP4 experts and higher-precision tensors | 69 KDA and 24 gated MLA layers; Kimi K3 license |
| DeepSeek-V4-Flash-0731 | Flash backbone: 284B, 13B active | 166.9 GB, FP4/FP8 and other tensors | Updated release includes DSpark; MIT |
| DeepSeek-V4-Pro-0813 | Pro backbone: 1.6T, 49B active | 892.7 GB, FP4/FP8 and other tensors | Updated release includes DSpark; MIT |
| Qwen3.8-27B | Dense 27B language model | 55.6 GB, BF16 | Vision and hybrid attention; Apache-2.0 |
| Qwen3.8-2.4T-A95B | 2.4T, 95B active | 4.892 TB, BF16 | Text checkpoint; 69 linear and 23 full-attention layers; model-specific license |
| MiniMax-M3 | About 428B, 23B active | 869.2 GB, BF16/F32 | Native multimodality and MiniMax Sparse Attention; inspect its license |
GLM-5.3-Flash is a different hybrid architecture, not merely a smaller GLM-5.3. Qwen's hosted Max product advertises capabilities beyond the downloadable text checkpoint. MiniMax's separately named quantized repositories are different artifacts from its BF16 release. Re-check the exact repository before reusing a sizing result.
The DeepSeek suffix matters too: 0731 and 0813 supersede previews and attach speculative components. Their larger checkpoint counts do not mean every target token activates all those extra parameters. The original DeepSeek V4 card distinguishes FP8 base checkpoints from mixed FP4/FP8 post-trained checkpoints. Reading only quant_method: fp8 misses expert_dtype: fp4. See the original release table.
Two trillion-scale models can need very different memory
Decimal terabytes
1.561 TB
4.892 TB
At MXFP4's ideal packing, 32 four-bit values plus one eight-bit scale occupy 17 bytes: 17 / 32 = 0.53125 bytes per value. That applies to tensors stored in that format, not every tensor in a model. Kimi's configuration excludes attention, shared experts, and other components from this quantization. Its actual index is stronger evidence than assuming an expert percentage.
For a hypothetical 2,304 GB aggregate GPU budget, 1,560.9 GB of stored Kimi tensors leaves 743.1 GB by subtraction. That is an optimistic aggregate remainder, not an available KV pool. Replication, padding, conversion, recurrent state, graphs, and temporary buffers reduce it; one rank can overflow while others have room. Weight Formats: FP8 Blocks, MXFP4 and AWQ explains how to account for these formats.
Sparsity changes work, but batches change reuse
Kimi's published active fraction is 104 / 2,800 = 3.7%. For Qwen's large model it is 95 / 2,400 = 4.0%. Keep every required expert resident somewhere in a resident-weight deployment, even though one token visits a subset. Offloading changes the memory constraint by introducing transfers.
Do not turn active parameters directly into HBM bytes per decode step. Several tokens can reuse an expert's weights; diverse routing can touch many experts. Attention reads, communication, precision, and load imbalance also matter. Expert Parallelism addresses placement; a measured concurrency sweep determines whether the resulting service meets its latency target.
Attention names describe different state
| Architecture example | State to budget | Common mistaken shortcut |
|---|---|---|
| GLM-5.3 latent attention | Compressed latent, rotary component, indexer state and engine overhead | Divide every cache tensor by tensor-parallel degree |
| Kimi K3; GLM-5.3-Flash | Growing attention cache plus recurrent state and retained prefix snapshots | Treat linear layers as having no memory |
| Qwen3.8-27B | 16 full-attention layers plus 48 linear layers | Apply the full-attention formula to all 64 layers |
| DeepSeek V4; MiniMax M3 | Architecture-specific compression, selection and auxiliary state | Assume sparse selection means old tokens can all be discarded |
From Qwen's configuration, the BF16 full-attention component alone is 2 × 16 layers × 4 KV heads × 256 × 2 bytes = 65,536 bytes, or 64 KiB per token. At 32,768 positions that is 2 GiB per sequence before recurrent state and replication. Its native context limit is 262,144; extending context is an additional configuration and validation decision.
From GLM's configuration, a BF16 compressed-latent representation over 78 main layers costs (512 + 64) × 2 × 78 = 89,856 bytes, or 87.75 KiB per token, before indexer and other state. At 1,048,576 positions this component is 87.75 GiB, or 94.22 GB. A GPU holding that cache still needs its weights and runtime buffers. Multi-Head Latent Attention and Sparse Indexers explains the distinction between retained state and selected attention work.
What to bring to a deployment review
Bring the exact revision, license, supported engine build, per-rank allocation report, and a workload with prompt/output distributions. Test chat encoding, reasoning controls, tool parsing, multimodal inputs where supported, and long-context quality. A repository creation timestamp is not proof of a public release date; this page dates its verification instead.
Choose the smaller model when it passes your task evaluation and offers better cost at the required latency. Reverse that decision when the larger model's measured quality improvement justifies its idle capacity and operational cost. Neither the newest release nor the highest benchmark score answers that question alone.
Key takeaways
- Kimi's 1.561 TB and Qwen's 4.892 TB checkpoints demonstrate why total parameters need a released format attached.
- Qwen3.8-27B is dense; the 2026 open-weight landscape is not exclusively sparse.
- GLM's 87.75 KiB and Qwen's 64 KiB examples cover specified attention components, not complete per-GPU allocations.
- Distinguish model revisions, API products, checkpoint bytes, and measured serving capacity before recommending hardware.