Will AI models with public weights become more efficient at the same pace as closed, proprietary ones — or will one group pull ahead?
Will open-weight and closed-weight algorithmic efficiency trends diverge in the future?
This explores whether the cost of reaching a given level of AI capability (algorithmic efficiency) will fall at different rates for models whose weights are publicly released than for models kept proprietary.
This explores whether the cost of reaching a given level of AI capability (algorithmic efficiency) will fall at different rates for open-weight models and closed models. The collection has no data that tracks efficiency trends over time for either group, so it can't answer the forecast directly. What it does offer are clues about how efficiency gains spread, and those clues point more toward convergence than divergence.
The first clue is that many recent efficiency gains are simple, published, and easy to copy. A systematic re-test of reinforcement-learning methods found that two basic tricks, advantage normalization and token-level loss aggregation, let plain PPO match more complex algorithms like GRPO and DAPO Can two simple techniques match complex RL algorithms?. Architecture-level ideas follow the same pattern. Running one transformer block twice instead of fetching new weights saves time on phones Does recomputing weights cost less than moving them on mobile?. Pairing cheap N-gram lookup memory with Mixture-of-Experts beats pure MoE at the same parameter and compute budget Can lookup memory and computation work together better than either alone?. Once a trick like these appears in a paper, anyone with weights can use it. That makes a lasting algorithmic lead hard to keep.
The second clue is that the gap may come from training rather than algorithms. The same PPO study found that the pretrained base model, not the choice of algorithm, sets the performance ceiling. A separate line of work shows that reasoning models stay ahead of non-reasoning models however much inference compute the non-reasoning models get, because training teaches a way of using extra tokens productively Can non-reasoning models catch up with more compute?. If the two groups do pull apart, the cause may be pretraining data and training recipes rather than clever algorithms. That would look like an efficiency gap without actually being one.
The third clue concerns the market. A study of over 100 trillion tokens on OpenRouter found that open-weight models such as DeepSeek and Qwen handle more than 30% of traffic, even in areas where closed models lead on capability How much do open-weight models actually get used in practice?. Users trade raw capability against cost and speed. This gives open-weight developers a strong reason to compete on efficiency in particular. Open and closed labs may end up optimizing for different goals, cost per token versus peak capability, even when they draw on the same ideas.
The last clue is that some limits apply to every model built the same way. Autoregressive LLMs level off at about 55–60% constraint satisfaction on optimization problems regardless of scale Do larger language models solve constrained optimization better?, because generating one token after another gives them no way to take back an earlier step Why does autoregressive generation fail at constraint satisfaction?. A wall like this blocks open and closed models alike, and it tends to pull trends together until someone changes the architecture.
Sources 7 notes
Advantage normalization and token-level loss aggregation allow critic-free PPO to surpass more complex algorithms. Systematic evaluation shows most RL techniques are setup-sensitive; the pretrained prior, not algorithm choice, sets performance ceiling.
MobileLLM shows that on memory-bound mobile hardware, sharing weights between adjacent transformer blocks by recomputing one block twice uses less latency than fetching separate weights, gaining accuracy with no parameter increase.
Engram combines O(1) N-gram lookup with Mixture-of-Experts routing, revealing a U-shaped scaling law where balanced allocation to both mechanisms outperforms either alone. Gains appear largest in reasoning and code rather than pure retrieval.
Reasoning models persistently outperform non-reasoning models regardless of inference budget because training instills a reasoning protocol that makes additional tokens productive. The gap is fundamentally about deployment mechanisms and training structure, not raw capability.
Analysis of 100+ trillion tokens on OpenRouter found open models like DeepSeek and Qwen captured large usage shares even where closed models lead on capability. Users trade cost, latency, and other factors against capability, creating persistent market heterogeneity.
Show all 7 sources
Across constrained-optimization tasks, LLMs converge to ~55–60% constraint satisfaction independent of architecture, parameter count, or training regime. Reasoning models do not systematically outperform standard models, suggesting a fundamental ceiling rather than a scaling gap.
The performance ceiling on constraint satisfaction problems is not a model-quality issue but an architectural limitation: autoregressive transformers cannot retract emitted tokens, while CSP solvers fundamentally depend on discarding invalid partial assignments. Symbolic solver integration works because it supplies what the architecture lacks.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- ZebraLogic: On the Scaling Limits of LLMs for Logical Reasoning
- Can Large Language Models Reason and Optimize Under Constraints?
- Single-Agent LLMs Outperform Multi-Agent Systems on Multi-Hop Reasoning Under Equal Thinking Token Budgets
- A Survey on Test-Time Scaling in Large Language Models: What, How, Where, and How Well?
- MiniMax-M1: Scaling Test-Time Compute Efficiently with Lightning Attention
- MobileLLM: Optimizing Sub-billion Parameter Language Models for On-Device Use Cases
- Conditional Memory via Scalable Lookup: A New Axis of Sparsity for Large Language Models
- Reasoning Models Can Be Effective Without Thinking