The next useful direction is a question whose answer would change our understanding. This page turns the model studies into a research agenda: what to change, what to hold fixed, what to measure, and what could falsify the explanation. It is not a forecast or a claim that these questions are unsolved everywhere.
An idea is not yet a conclusion
A comparison with a shared contract
Baseline and candidate receive matched data, tasks and a declared training or active-compute budget. Runtime comparisons also identify kernel, precision and hardware.
A result with visible tradeoffs
Read quality, latency, retained state, total weights and communication separately. Then inspect failure slices and repeat under changed conditions.
An experimental design proposed here, not an experiment or benchmark result.
Figure 1. A shared contract, several outcomes, and visible failure cases. Experimental design proposed here; no measurements are shown.
Use Figure 1 to define the comparison before reading the result. Architecture, data, optimizer, precision and policy learning often change together. A full model result establishes what that system achieved under its evaluation conditions. A controlled ablation is better evidence for which ingredient produced a particular effect.
Questions worth testing
| Direction | One controlled change | What would make it useful? | Start with |
|---|---|---|---|
| Memory | Change recurrent/global ratio at a declared budget | Retrieval under distance and distractors, plus state and decoding cost | Kimi Linear, MiniMax 01 |
| Selected history | Change selector or block budget | Preserve tasks requiring several distant facts; count index overhead and cache | GLM-5, MiniMax M3 |
| Capacity | Change expert width, count or sharing | Distinguish active work, total weights, communication and quality | Mixtral, DeepSeek-V3 |
| Depth | Change residual reads, writes or source granularity | Improve final behavior without unstable optimization or excessive traffic | Kimi K3, Qwen3.8-Next |
| Visual evidence | Change resolution, merging or entry depth | Gains at matched visual and training budgets, including fine detail and conflicting evidence | Qwen3-VL, Gemma 3 |
| Position and reach | Change scaling or sampling while controlling evidence | Preserve use of the right region, frame or distant fact, not only accepted length | Qwen2.5 1M, Qwen2.5-VL |
These are proposed comparisons. A matched active compute experiment and a matched total parameter experiment answer different questions; choose and state the contract rather than claim both were controlled.
Negative results are part of the map
Qwen3.8-Next reports that a residual read shortcut can look acceptable in pretraining metrics yet degrade later quality.1 MiniMax’s sparse attention report likewise shows that aggregate competitiveness does not mean every long context subtask improves.2
Such evidence should narrow a claim. “Useful in this setup” is stronger scientifically than an unsupported “better in general.” A failed replication also needs its setup checked before declaring the mechanism false.
What would change these pages?
A new study earns a place when it supplies a distinct mechanism, a useful combination, an inspectable implementation, or counterevidence to an existing explanation. Record the baseline, versions, task, budget and failure slices beside the result. Connect it to the relevant mechanism page.
For a proprietary model, public disclosure and a plausible reconstruction remain different evidence. An unknown internal detail must not become fact by analogy with a released model.
Selected examples rechecked October 10, 2026. No controlled experiments were performed here. Return to the design map or the model catalog to follow a specific question.
References
-
Qwen Team. On the Design of Qwen3.8-Next Architecture, 2026, section 2.2 and later-stage residual read ablations. ↩
-
MiniMax. MiniMax Sparse Attention, 2026, long context evaluation and continued pretraining comparisons. ↩