AI Models

MTP speculative decoding — acceptance length

Ablation on coding scenarios; +20% over baseline.

← Back to GLM-5.2 Benchmarks

About this data

GLM-5.2’s multi-token-prediction layer for speculative decoding is improved step by step: IndexShare + KVShare, then rejection sampling, then an end-to-end TV loss lift the acceptance length from 4.56 to 5.47 — a 20% gain that directly speeds up inference. This is an internal ablation, not a cross-model comparison.

MTP speculative decoding — acceptance length

Ablation on coding scenarios. Each technique lifts acceptance length; +20% over baseline.

View data & sources →

Data table

MTP speculative decoding — acceptance length — mtp_ablation data table (GLM-5.2 Benchmarks)
step series acc_len source_ref value_basis
Baseline mtp_ablation 4.56 zai-glm52 MTP ablation table — baseline 4.56
+ IndexShare + KVShare mtp_ablation 5.1 zai-glm52 MTP ablation table — 5.10
+ Rejection Sampling mtp_ablation 5.29 zai-glm52 MTP ablation table — 5.29
+ End-to-end TV Loss mtp_ablation 5.47 zai-glm52 MTP ablation table — 5.47 (+20% over baseline)

Methodology & sources

Last updated: Jul 17, 2026

Methodology

Source-backed values for all six charts come from the Z.ai GLM-5.2 launch post (2026-06-16): long-horizon coding (FrontierSWE, PostTrainBench, SWE-Marathon), standard coding (Terminal-Bench 2.1, SWE-bench Pro, NL2Repo, DeepSWE, ProgramBench), reasoning (HLE, CritPt, AIME, HMMT, GPQA-Diamond, IMOAnswerBench), agentic (MCP-Atlas, Tool-Decathlon), the GLM-5.2 vs GLM-5.1 generational leap, and the MTP acceptance-length ablation. Every numeric point carries a sources[].ref and a value_basis naming the table row. Scores are reproduced verbatim from the Full Benchmark Table; cells the vendor published as "-" (not reported) are omitted, never estimated. CAVEAT: these are the model developer’s self-reported figures under their stated harnesses and prompts (see the post’s footnotes); some competitor HLE/CritPt cells are full-set scores. They are comparison-as-published, not an independent eval. Architecture context (not charted): 1M-token context (up from 200K), IndexShare cuts per-token indexer FLOPs ~2.9× at 1M, MIT-licensed open weights. Re-verified 2026-06-22.

Sources

Comparisons are informative, not definitive. See each source for definitions and limits.

Continue researching

How relevant was this information?