7arXiv cs.CL (Computation and Language)·
Mach-Mind-4-Flash: 35B MoE agentic model matching 100B-class performance via post-training optimization
Mach-Mind-4-Flash is a 35B-parameter Mixture-of-Experts model with only 3B activated parameters that achieves performance comparable to 100B-class models through post-training techniques alone. The pipeline combines a unified RL/OPD training infrastructure with multi-teacher scheduling, parallel domain-specific RL experts fused via Multi-Teacher On-Policy Distillation (MOPD), and Hybrid Median-length Policy Optimization (HMPO) which compresses reasoning chains 19-46% with minimal accuracy loss. Benchmark results include 92.70 on AIME'26, 82.82 on IFBench, and 75.80 on BFCL-v4, claiming to lead or match models 10-30x its activated size at a fraction of inference cost. The work is notable for demonstrating that post-training optimization can close large gaps in activated parameter count for agentic tasks.
Related guides (3)
Related events (8)
Agents-A1: 35B MoE agent matches trillion-parameter models via horizon scaling
Researchers introduce Agents-A1, a 35B Mixture-of-Experts model that claims to match or exceed trillion-parameter models like Kimi-K2 and DeepSeek V4 on long-horizon agentic benchmarks. The approach scales agent trajectory length (averaging 45K tokens) and heterogeneous agent abilities rather than raw parameter count, using a three-stage training recipe including multi-teacher domain-routed distillation. On benchmarks such as SEAL-0, IFBench, HiPhO, and FrontierScience-Olympiad, Agents-A1 achieves leading or competitive results against models with roughly 30x more parameters. The work proposes a practical efficiency path for agentic capability scaling without proportional compute scaling.
MOPD: Multi-Teacher On-Policy Distillation for integrating multiple RL-trained capabilities in LLMs
Researchers propose Multi-teacher On-Policy Distillation (MOPD), a post-training paradigm that first trains domain-specialized RL teacher models, then distills them into a student model using on-policy rollouts to eliminate exposure bias. Evaluated on Qwen3-30B-A3B, MOPD outperforms Mix-RL, Cascade RL, Off-Policy Finetune, and Param-Merge baselines while preserving nearly all per-domain capability. The method has been deployed in production for MiMo-V2-Flash, an industrial-scale frontier model, validating its practical applicability. The approach also enables parallel, decoupled development of domain teachers, reducing cross-domain interference in multi-capability post-training.
ZEDA: Post-Trained MoE Models Can Skip Half Their Experts via Self-Distillation
This paper introduces Zero-Expert Self-Distillation Adaptation (ZEDA), a framework that converts static post-trained Mixture-of-Experts (MoE) language models into dynamic ones without pre-training from scratch. ZEDA injects parameter-free zero-output experts into each MoE layer and uses two-stage self-distillation with the original model as a frozen teacher. Applied to Qwen3-30B-A3B and GLM-4.7-Flash across 11 benchmarks, ZEDA eliminates over 50% of expert FLOPs with marginal accuracy loss and achieves approximately 1.20× end-to-end inference speedup, outperforming the strongest dynamic MoE baseline by 4–6 points.
MobileMoE: Scaling Mixture-of-Experts for Sub-Billion Parameter On-Device Deployment
MobileMoE introduces a family of on-device MoE language models with 0.3–0.9B active parameters and 1.3–5.3B total parameters, targeting mobile deployment under memory and compute constraints. The authors derive an on-device MoE scaling law identifying a sweet spot of moderate sparsity with fine-grained and shared experts, then train models through a four-stage recipe including quantization-aware training on open-source data. Across 14 benchmarks, MobileMoE matches or exceeds leading dense on-device LLMs with 2–4× fewer inference FLOPs, and delivers 1.8–3.8× faster prefill and 2.2–3.4× faster decode than dense baselines on commodity smartphones at comparable INT4 memory.
AXPO: Agent Explorative Policy Optimization Addresses Thinking-Acting Gap in Multimodal Agentic Reasoning
This paper identifies a structural asymmetry in agentic reasoning called the 'Thinking-Acting Gap,' where tool use is attempted in only ~30% of rollouts under standard RL training (GRPO), and all-wrong tool-using subgroups suppress learning signals. The authors propose AXPO (Agent eXplorative Policy Optimization), which fixes the thinking prefix and resamples tool calls for all-wrong subgroups, combined with uncertainty-based prefix selection. Evaluated across nine multimodal benchmarks on Qwen3-VL-Thinking at multiple scales, SFT+AXPO outperforms SFT+GRPO by +1.8pp on both Pass@1 and Pass@4 at 8B, with the 8B model surpassing the 32B baseline on Pass@4 using 4× fewer parameters.
Mixtral 8x7B: Mistral AI Releases Sparse Mixture-of-Experts Open-Weight Model
Mistral AI has released Mixtral 8x7B, a sparse mixture-of-experts (SMoE) model with 46.7B total parameters but only 12.9B active parameters per token, enabling inference speed and cost equivalent to a 12.9B model. Licensed under Apache 2.0, Mixtral outperforms Llama 2 70B on most benchmarks and matches or exceeds GPT-3.5, with support for 32k context, five European languages, and strong code generation. An instruction-tuned variant (Mixtral 8x7B Instruct) achieves 8.3 on MT-Bench, claimed best among open-source models at release. The model is deployed behind Mistral's mistral-small API endpoint and supported via vLLM with Megablocks CUDA kernels.
Z.ai releases GLM-5.2, a 753B MoE open-weights model claiming top open-model ranking on agentic coding benchmarks
Z.ai released GLM-5.2, a 753-billion-parameter mixture-of-experts open-weights model optimized for long-running agentic coding tasks, with a 1-million-token input context and MIT license. The model ranks first among open-weights models on Artificial Analysis's Intelligence Index v4.1 (score 51, behind Claude Opus 4.8 at 56 and GPT-5.5 at 55) and leads all models on PostTrainBench, a benchmark for agentic fine-tuning tasks. Key technical contributions include a modified sparse attention indexer applied every four layers (cutting per-token computation 2.9x at 1M context), a switch from GRPO to PPO for long-horizon RL training, and a reward-hacking mitigation pipeline using rule-based filters and a judge model. API pricing is substantially below comparable proprietary models, and the release coincides with U.S. government restrictions on access to Anthropic's frontier models.
Mistral Small 4: Unified Multimodal, Reasoning, and Coding MoE Model Released Under Apache 2.0
Mistral AI has released Mistral Small 4, a 119B-parameter Mixture-of-Experts model (6B active per token) that unifies capabilities previously split across Magistral (reasoning), Pixtral (multimodal), and Devstral (coding agents) into a single open-weights model. The model features a 256k context window, configurable reasoning effort via a `reasoning_effort` parameter, native text and image input support, and is released under Apache 2.0. Mistral claims 40% latency reduction and 3x throughput improvement over Mistral Small 3, with benchmark results showing competitive performance against GPT-OSS 120B and Qwen models while producing significantly shorter outputs. The release includes day-0 availability as an NVIDIA NIM and support across vLLM, llama.cpp, SGLang, and Transformers.