PublishersMarkTechPost

MarkTechPost

6 documents read from this publisher.

Z.ai Releases GLM-5.3-Flash: A 320B-A18B Natively Multimodal MoE With a 1M-Token Context

news analysis · news · 2026-08-26 · 7 techniques

Mixture of Experts Hybrid Attention Multi-Token Prediction FP8 Manifold-Constrained Hyper-Connections Encoder-Prefill-Decode disaggregation IndexPool

Z.ai Ships GLM-5.3 Without Retraining the Base Model: Better at Complex Coding and Long-Horizon Tasks

news analysis · news · 2026-08-14 · 4 techniques

Vulnerability-discovery data inclusion post-training scaling private benchmark for evaluation scaling post-training with more task environments, environment types, and longer training

Poolside Releases Laguna S 2.1, an Open-Weight Agentic Coding Model Punching Above Its Weight Class on SWE-Bench Multilingual

news analysis · news · 2026-07-21 · 8 techniques

Mixture of Experts Sparse expert activation Maximum thinking effort FP8-precision reinforcement learning Multi-precision weight formats (BF16/FP8/INT4/NVFP4)Publication of unedited evaluation trajectories Thinking modes (off and max)TRT-LLM inference optimization to NVFP4 on Blackwell

Tencent Releases Hy3: An Open 295B Mixture-of-Experts (MoE) Model with 21B Active Parameters and 256K Context

news analysis · news · 2026-07-06 · 13 techniques

Mixture of Experts Multi-Token Prediction Speculative decoding FP8 Chain-of-thought reasoning Top-8 expert routing Reasoning parser Speculative sampling AngelSlim Generic tool-call parser Anti-hallucination training Joint SFT and RL Low-bit quantization

StepFun Releases Step 3.7 Flash: A 198B MoE Vision-Language Model for Coding Agents and Search Workflows

news analysis · news · 2026-05-29 · 11 techniques

Mixture of Experts Configurable reasoning effort FP8 BF16 NVFP4 quantization Advisor strategy Python tool GGUF Visual Search Tool Apache 2.0 License Emergent Compositional Tool Use

Xiaomi Releases MiMo-V2.5-Pro and MiMo-V2.5: Matching Frontier Model Benchmarks at Significantly Lower Token Cost

news analysis · news · 2026-04-22 · 9 techniques

1M-token context window OpenCode Claude Code Native multimodal understanding agentic model harness awareness Kilo Native visual and audio understanding ngspice simulation