Four Labs, One Recipe: Open-Weight Frontier Models Converged on 3-4% Active Params and Hybrid Attention in 35 Days
In 35 days between late July and late August 2026, four separate labs released frontier-class open-weight models. When you open the config files side by side, the architecture is almost the same.
The four models:
| Model | Lab | Release | Total Params | Active Params | Active % |
|---|---|---|---|---|---|
| Kimi K3 | Moonshot | July 27 | 2.78T | 104B | 3.7% |
| Qwen3.8-2.4T | Alibaba | August 12 | 2.4T | 95B | 4.0% |
| DeepSeek V4-Pro | DeepSeek | August 13 (GA) | 1.6T | 49B | 3.1% |
| GLM-5.3-Flash | Z.ai | August 26 | 320B | 18B | 5.6% |
The Shared Recipe
All four converge on the same three-part design:
Sparse activation. Store trillions of parameters; activate 3-6% per token. Capacity lives in total parameter count. Intelligence lives in active parameter count. The key engineering insight is that these two objectives can be decoupled.
Hybrid attention. Run approximately one exact-attention (full MLA or gated MLA) layer for every three cheap linear-attention layers (Kimi Delta Attention, Gated DeltaNet, Compressed-Sparse Attention, KDA). Exact attention is expensive but expressive; linear attention is fast but weaker on long-range dependencies. The ratio varies slightly by lab — Kimi uses 24 gated MLA out of 93 layers; GLM-5.3-Flash uses 11 DeepSeek Sparse Attention out of 45 — but the hybrid principle is identical.
Long context as a baseline. All four support 1M tokens natively or near-natively. Kimi K3 and DeepSeek V4-Pro are confirmed at 1M; Qwen3.8 ships at 262K with the model card noting extensibility to roughly 1M.
The additional shared choices: Muon optimizer, multi-token prediction, 4-bit expert quantization in production. These are all independent design decisions that each lab apparently arrived at separately, or copied very quickly.
Where They Stand on Intelligence
Artificial Analysis runs its own independent v4.1.1 Intelligence Index across nine evaluations, separately from vendor-reported benchmarks. As of late August 2026:
- Closed frontier (Claude Opus 5, Fable 5, GPT-5.6 Sol, Grok 4.6): 61-63
- Best open weights (Kimi K3 max, GLM-5.3 max, Qwen3.8): 57-60
A 1-3 point gap on the AA index. The open frontier is not matching the closed frontier on this measure, but it is close enough that the gap is practical rather than categorical.
The Benchmark Caveat
Vendor-reported agentic coding scores on Terminal-Bench 2.1 cluster tightly: GPT-5.6 Sol 88.8, Kimi K3 88.3, GLM-5.3 88.2, DeepSeek V4-Pro 87.9, Qwen3.8-Max 86.6. These numbers are widely cited but uniformly self-reported on differing harnesses and effort settings. llm-stats tracks Terminal-Bench 2.1 at 0 verified / 30 self-reported as of late August.
The directional message is real: open models reached the frontier pack on agent-relevant tasks. The exact ordering should wait for third-party harnesses.
Cost Position
The Artificial Analysis Intelligence Index gap is 1-3 points. The cost gap is substantially larger in the other direction. Kimi K3 open-weight deployment runs at roughly one-third the per-task cost of Claude Opus 5 or GPT-5.6 Sol on LiveBench’s cost-per-successful-task metric. At $0.348 per successful task for Kimi K3 open versus $0.699 for Claude Opus 5 Thinking, the cost-adjusted argument for open weights has become concrete.
What Closed Labs Did This Summer
The closed frontier is not standing still. Claude Opus 5 hit 97.0% SWE-bench Verified. GPT-5.6 Sol reached 96.2%. Grok 4.6 hit 95.6%. These scores are from independent neutral harnesses, not vendor submissions.
The architecture convergence is not a signal that open weights have surpassed closed. It is a signal that the recipe for building a frontier-adjacent model is now known, public, and replicable. Four labs proved it in 35 days.