Nishi Compare -- Frontier Radar: LLM and Inference Stack

Where the 2025/26 research frontier is moving, and where Nishi should test, deepen, or research next.

How this works. The nishi researcher banked the current (2024-2026) frontier for this domain over sovereign TLS (keyless OpenAlex/arXiv). For each capability axis this radar measures frontier momentum = how often the frontier corpus mentions it, times our organ presence (probed on disk). Absent + hot = a research opportunity / where to test. Corpus measured: 1774751 bytes. Generated by nx_swcompare_gapmap over a data-only domain spec (llm.q + llm.axes) — zero domain-specific code.

Capability axis	Nishi organ	Frontier momentum (2025/26)	Signal
Attention	present	40 mentions	COVERED
Continuous-batching serving	present	35 mentions	COVERED
Int4 quantized matmul	present	25 mentions	COVERED
Autodiff / training	present	21 mentions	COVERED
Transformer forward (no-float)	present	21 mentions	COVERED
Test-time compute / reasoning	present	11 mentions	COVERED
SIMD integer GEMM	present	5 mentions	COVERED
KV-cache / paged attention	present	3 mentions	COVERED
Speculative decoding	present	3 mentions	COVERED
Mixture of experts (sparse)	present	3 mentions	COVERED
Fused dequant GEMM (Marlin-class)	present	2 mentions	COVERED
LayerNorm / normalization	present	1 mentions	COVERED
BPE tokenizer	present	1 mentions	COVERED
Long-context inference	absent	0 mentions	dormant

frontier momentum is a mechanical keyword count over real 2024-2026 works; our presence is probed on disk. A discovery signal, not a ranking of importance. Zero JS, zero trackers.

Frontier Radar — LLM and Inference Stack