Baidu has released ERNIE 5.5 Flash Base, a 280-billion-parameter mixture-of-experts foundation model. It was pre-trained on 33.5 trillion tokens. Baidu shipped the model without fine-tuning so developers can adapt it for their own uses.
This paper reveals that expert-reduction methods for MoE LLM serving, which speed up inference by running fewer experts per token, can weaken refusal of harmful requests. The authors introduce SafeStart, which runs all experts for the prompt and first eight generated tokens, then reduces experts afterward, preserving most speedup while restoring safety. In vLLM on an RTX A6000, SafeStart keeps at least 85% of the offline decode speedup at batch sizes 8 and 32 and reduces harmful compliance increases to near baseline levels across three models.
Why it matters: Provides a concrete, recent finding that expert reduction in MoE serving has safety costs and offers a practical mitigation, valuable for AI serving systems and safety research.
NVIDIA's technical blog details three optimizations for mixture-of-experts (MoE) inference on a single DGX Spark desktop system, using Ant Group's Ling-3.0-flash model. The optimizations—Humming MXFP4 MoE kernels, an FP8 LM head, and built-in MTP speculative decoding—boosted GSM8K decode throughput by 42.5% (from 37.81 to 53.89 tokens/s) with no statistically significant accuracy change. The post also highlights the memory-bound nature of MoE decoding and the integration trap when stacking FP8 LM head with MTP, requiring a shared LM-head module.
Why it matters: Provides concrete, actionable performance optimizations for running large MoE models on edge hardware, with real benchmark results and practical integration guidance.
Mistral Large 4 (ML4) is a new multimodal mixture-of-experts model with 52B active parameters out of 1T total, optimized for reasoning and long-horizon agentic tasks. The model is built using a scalable agent-based approach, though specific technical details are limited in the provided text. This release signals a significant advancement in large-scale MoE architectures for complex AI applications.
Why it matters: The announcement of a major new model from Mistral with a novel architecture and focus on agentic tasks is highly relevant to AI practitioners and researchers.
A new paper combines calibrated multi-teacher on-policy distillation with execution optimizations for dynamically routed mixture-of-experts models. On Turing-20B-A2B, the authors report retaining at least 90.16% of measured specialist gains in every evaluated domain and outperforming shared supervised fine-tuning and Qwen3-8B on four reasoning-enabled mathematics and code benchmarks at a matched response cap. A PPU-based profile shows consolidation step time falling from 398.7 to 139.3 seconds, with the results attributed to the complete training and execution framework rather than a single component.
Why it matters: The study offers concrete evaluation and execution results for researchers seeking to combine specialist models while reducing MoE post-training costs.
The paper proposes Domain-Specialized Upcycling (DS-Upcycling), which expands a dense Transformer layer with additional experts and a router rather than attaching per-domain LoRA adapters. A new domain-aware supervision signal trains the router to select the right expert for each input domain, and fine-tuning on a new domain updates only its matching expert and the router while freezing everything else. Across adapted domains the method reports the highest average performance among baselines while preserving general capability. With only 256 examples per domain for router supervision, routing accuracy after adaptation reaches 90-98%, versus 43-70% for routers trained without the proposed approach.
Prime Intellect released Prime Mega MoE, a training-capable fused MoE kernel built on DeepSeek's DeepGEMM Mega MoE, adding an overlapping backward pass. It is available as a dispatch backend in prime-rl and supports BF16, blockwise FP8 on Hopper, and MXFP8 on Blackwell. Prime Intellect reports 32–34% faster full training steps than its prime-rl baseline and 48–55% faster than TorchTitan on Qwen3 MoE models, with up to 3.11x forward-pass speedup.
Why it matters: MoE training teams get a concrete open-source kernel option that may cut GPU hours, with clear implementation details and reported caveats on where gains are largest.
Agent-Conditioned Routing (ACR) runs parallel trajectories inside one frozen sparse mixture-of-experts model by applying persistent, untrained shared and private expert-eligibility masks while keeping the learned, input-dependent token router. On a screened 126-question mathematics and physics benchmark, ACR yielded about twice as many problems with at least two distinct primary methods among soft-correct solutions as the strongest temperature setting on GLM 5.3 (37.3 versus 18.7) and 1.67 times as many on GPT-5.6 Sol (45 versus 27), at matched output budgets. A blinded four-expert audit of 30 problems also found more multi-method solutions with ACR on both backbones, though correctness was lower. ACR adds about 160 KiB of mask state, and a three-problem microbenchmark stayed within 3% of batched temperature sampling on latency and peak VRAM.
MiniMax has launched MiniMax M3.1 Flash on its API, a multimodal mixture-of-experts model with 428B total parameters, about 23B active, and a 1,000,000-token context window. It is priced at a flat $0.15 per 1M input tokens and $0.60 per 1M output tokens, half of M3's standard per-token rate. MiniMax reports near-frontier agentic results, including a top Automation-Bench v1.0.6 score of 62.2, but weaker hard-coding results such as 38.1 on Terminal Bench 4.0. Weights are due October 23, 2026, with
Why it matters: It gives developers concrete pricing, context, benchmark and deployment details for a newly released low-cost agentic model, with vendor-reported figures clearly flagged.
StepFun has published open FP8 and BF16 checkpoints for Step 5 Preview, a sparse Mixture-of-Experts model with 600B total parameters, 27B active per token and up to 1M tokens of context, targeting long-horizon agentic work. The release lists data-center-only self-hosting requirements, including 4x B300 or 8x H200 GPUs for the FP8 checkpoint, and an API priced at $1.00 per 1M input tokens and $2.70 per 1M output tokens. StepFun-reported results put it near GLM-5.3 and Kimi K3 on DeepSWE v1.1, whi
Why it matters: It gives developers a concrete new self-hostable open-weight option for agentic and long-context work, with clear hardware requirements, pricing and benchmark caveats.
Liquid AI released LFM2.5-24B-A2B, an open-weight 24B-parameter Mixture-of-Experts reasoning model with about 2B active parameters and 131,072-token context, targeted at local agent loops. It runs via GGUF/MLX on laptops and mini PCs, with cited speeds of 184 tokens/s on Apple M5 Max and 114 tokens/s on AMD Ryzen AI Max+ 395, plus single NVIDIA B300 serving. Liquid AI reports it leads or matches larger MoEs on several agentic, tool-use and instruction-following benchmarks, while trailing Qwen3.6
Why it matters: Provides a concrete open-weight release with on-device performance, agentic benchmark comparisons, licensing and deployment details useful for evaluating local agent models.
AWS described three kernel-level optimizations for fused Mixture-of-Experts inference on NVIDIA Hopper: a Swap-AB transposition to improve WGMMA utilization, routing FP8 tensor-quantized MoE to the Triton backend, and eliminating Triton JIT recompilations with do_not_specialize annotations. The post reports up to 10% higher end-to-end throughput and up to 9% faster Time to First Token on H200. The optimizations were upstreamed to vLLM and are used in the AWS LMI v28 container on Amazon SageMaker
Why it matters: Provides practical, open-source GPU kernel techniques and measured latency and throughput gains for teams deploying MoE models on Hopper and SageMaker.
Cloudflare released Clef-omni, an open-weight Apache-2.0 decision model that scores text, images, audio and video in a single call. It is post-trained from Qwen3-Omni-30B-A3B-Instruct as a 30B-total/3B-active MoE with 64,000-token context on Workers AI, priced at $0.15 per million tokens. Cloudflare reports strong intent classification results (97.7 macro-F1 on CLINC150+OOS) but weaker scores than text-focused Clef on reasoning and workflow evals, and Cloudflare's benchmark numbers were not inde
Why it matters: It documents a concrete new open-weight multimodal model release with architecture, pricing, latency and benchmark details useful for developers building multimodal agent pipelines.
Nous Research Unveils Token Superposition Training, Cutting Pretraining Time by 2–3x Amid Convergent Research ControversyNous Research has introduced a new large language model pretraining method called Token Superposition Training (TST), which short<!--autointro-->
NVIDIA officially launched its new multimodal reasoning model, Nemotron 3 Nano Omni, yesterday. It deeply integrates text, vision, and speech capabilities into a single model system and is currently available for free.As the latest member of the Nemo<!--autointro-->
