Version 2

A cluster of efficiency results published June 9–16, 2026 challenges the assumption that AI capability and inference cost scale with parameter count. MiniMax Sparse Attention (MSA), deployed in M3, cuts attention compute 28.4x at 1-million-token contexts [3][1]. A separate architectural change removes KV projection matrices to halve cache size at 3.1% perplexity cost [6], and a 4B-parameter model outperforms a 671B model on formal theorem proving [7]. Two additions from June 16 extend the picture: Tensordyne announced logarithmic arithmetic inference chips claiming 17x tokens-per-watt and 13x higher throughput over NVIDIA Blackwell [9], and TokenPilot reports 61–87% agent cost reduction via memory-centric context management [8].

Efficiency gains are appearing simultaneously across hardware, model architecture, memory management, and agent systems — layers that are largely independent of one another. If these results hold up under independent testing, the combined effect on inference cost and deployment accessibility would be large without requiring larger models.

A cluster of efficiency results published between June 9 and 16, 2026 puts pressure on the assumption that AI capability and inference cost scale primarily with parameter count. The findings come from different parts of the stack — attention computation, KV memory design, training data geometry, hardware arithmetic, and agent context management — and are not products of a single coordinated research program.

The most extensively covered result comes from MiniMax, which released its M3 model alongside a paper describing MiniMax Sparse Attention (MSA) [1][2]. Full attention computes relevance between every pair of tokens in a sequence, making computation grow quadratically with context length. MSA selectively attends to a subset of tokens, cutting attention compute 28.4x at 1-million-token contexts and delivering 14.2x faster prefill and 7.6x faster decoding on H800 GPUs while mostly matching full-attention benchmark performance [3]. NVIDIA published a technical blog covering deployment of M3 on its accelerated infrastructure [4], and commentators noted that the efficiency gains make long-context inference practically viable on hardware that would otherwise be compute-bound [5].

Three other software-layer results extend the efficiency case further. Research circulated June 9 shows that removing separate key and value projection matrices — standard transformer attention components — halves KV cache size with a 3.1% perplexity penalty [6]; this is a structural model change rather than a serving optimization. Pythagoras-Prover, a 4-billion-parameter model, outperforms DeepSeek-Prover-V2-671B on the MiniF2F Pass@32 formal theorem-proving benchmark, with its developers attributing the win to improved training data geometry rather than architectural changes [7]. TokenPilot replaces prompt truncation with ingestion-aware compaction and lifecycle-aware eviction, reporting 61–87% cost reductions on PinchBench and Claw-Eval benchmarks [8].

The most recent addition is a hardware claim. Tensordyne announced an inference chip built around logarithmic arithmetic implemented in silicon rather than software, reporting 17x higher tokens per watt and 13x higher throughput than NVIDIA Blackwell [9]. If these figures survive independent testing, the approach would represent a different path to inference efficiency than the software and model-architecture methods above — changing the arithmetic substrate rather than the model structure or serving strategy. No independent replication of Tensordyne's claims exists yet.