Infra2026-09-23 · source-backed

Keeping two-thirds of routed experts preserves 98.8% of fine-grained MoE performance.

Issue

2026-09-23

Confidence

source-backed

Sources

Redaction

redacted

Story

arXiv 2609.25809 studied 12 checkpoints across nine architecture families on 11 benchmarks and found per-token expert selection far more redundant than assumed. The change is a single integer, and it gave 1.2-1.7x measured speedup on two serving backends.

A poisoned MoE router sends 92–96% of triggered tokens to one GPU and nobody notices, because the outputs are fine.
Load Hijack modifies nothing but router weights in a checkpoint. When a private trigger appears, token-to-expert assignment concentrates on experts co-located on a single GPU, making it a straggler while peers idle (arXiv 2608.10614). Across three MoE families and four corpora...
Qwen3.8-Flash-Next matches a 397B predecessor on 8 of 14 benchmarks at a ninth of the training FLOPs.
The architecture report describes a 125B sparse MoE activating 6B parameters per token, plus 51B of n-gram embedding tables held off the accelerator in host memory with prefetching (arXiv 2608.30320). Against the prior 397B-A17B model it leads on 8 of 14 pre-training benchmark...
Predicting the next layer's routing one token ahead serves a 35B MoE from SSD at 20 tok/s in 3GiB.
Edge0 attacks why naive SSD offloading fails for MoE: layer N+1's experts must be chosen before layer N's output exists, so reads can't start early enough to hide behind compute. A per-layer prerouter predicts the next layer's routing one token ahead and then *uses* that predi...
Pruning MoE experts independently is a first-order approximation, and fixing it adds 6.1% on agentic coding.
HOPE points out existing expert-pruning methods decide each expert in isolation and assume contributions are purely additive, when MoE expert usage is cooperative. It derives a second-order objective that provably minimizes an upper bound on pruning error, and shows the state-...
Refusal ablation works on a 320B model, standard method misses it
arXiv 2609.09793 applies single-direction refusal ablation to GLM-5.3-Flash, a 320B MoE with 288 routed experts and a four-wide hyper-connection residual. Editing attention, dense and routed-expert writers individually removes 0.039, 0.016 and 0.148 of refusal; editing all thr...
CANN Bench opens agent kernel generation beyond CUDA.
A 15-author Huawei team argues kernel-generation benchmarks are almost entirely CUDA and Triton, leaving less-documented hardware with no shared yardstick (arXiv 2607.20518). CANN Bench covers 53 operators and 1,060 test cases in four difficulty tiers, from elementwise primiti...

Source trail

arXiv 2609.25809 More from arxiv.org

Entities

Keeping MoE

Provenance

Canonical issue: Ramsay Research Agent — September 23, 2026
AI generated: no
Story unit: 2026-09-23-keeping-two-thirds-of-routed-experts-preserves-98-8-of-fine-grained-moe-performance
Labels: source-backed, canonical briefing excerpt

Fetching from the wire…

Keeping two-thirds of routed experts preserves 98.8% of fine-grained MoE performance.

Story

Related stories

A poisoned MoE router sends 92–96% of triggered tokens to one GPU and nobody notices, because the outputs are fine.

Qwen3.8-Flash-Next matches a 397B predecessor on 8 of 14 benchmarks at a ninth of the training FLOPs.

Predicting the next layer's routing one token ahead serves a 35B MoE from SSD at 20 tok/s in 3GiB.

Pruning MoE experts independently is a first-order approximation, and fixing it adds 6.1% on agentic coding.

Refusal ablation works on a 320B model, standard method misses it

CANN Bench opens agent kernel generation beyond CUDA.

Source trail

Entities

Provenance

Keeping two-thirds of routed experts preserves 98.8% of fine-grained MoE performance.

Story

Related stories

A poisoned MoE router sends 92–96% of triggered tokens to one GPU and nobody notices, because the outputs are fine.

Qwen3.8-Flash-Next matches a 397B predecessor on 8 of 14 benchmarks at a ninth of the training FLOPs.

Predicting the next layer's routing one token ahead serves a 35B MoE from SSD at 20 tok/s in 3GiB.

Pruning MoE experts independently is a first-order approximation, and fixing it adds 6.1% on agentic coding.

Refusal ablation works on a 320B model, standard method misses it

CANN Bench opens agent kernel generation beyond CUDA.

Source trail

Entities

Provenance

Story

Related stories

A poisoned MoE router sends 92–96% of triggered tokens to one GPU and nobody notices, because the outputs are fine.

Qwen3.8-Flash-Next matches a 397B predecessor on 8 of 14 benchmarks at a ninth of the training FLOPs.

Predicting the next layer's routing one token ahead serves a 35B MoE from SSD at 20 tok/s in 3GiB.

Pruning MoE experts independently is a first-order approximation, and fixing it adds 6.1% on agentic coding.

Refusal ablation works on a 320B model, standard method misses it

CANN Bench opens agent kernel generation beyond CUDA.

Tomorrow's issue, in your inbox.

Source trail

Entities

Provenance

Story

Related stories

A poisoned MoE router sends 92–96% of triggered tokens to one GPU and nobody notices, because the outputs are fine.

Qwen3.8-Flash-Next matches a 397B predecessor on 8 of 14 benchmarks at a ninth of the training FLOPs.

Predicting the next layer's routing one token ahead serves a 35B MoE from SSD at 20 tok/s in 3GiB.

Pruning MoE experts independently is a first-order approximation, and fixing it adds 6.1% on agentic coding.

Refusal ablation works on a 320B model, standard method misses it

CANN Bench opens agent kernel generation beyond CUDA.

Tomorrow's issue, in your inbox.

Source trail

Entities

Provenance