Releases·Model Release
China Telecom releases Xing4.0-29B-A4B MoE, trained on Ascend NPUs with 256K context
China Telecom's AI unit released Xing4.0-29B-A4B on Hugging Face, a 29B-parameter MoE (4B active) successor to the TeleChat series, natively supporting 256K context. It is billed as the first model of this scale trained entirely on Ascend NPUs with MindSpore, with agent-oriented architecture and open ecosystem compatibility for fine-tuning and deployment.
169 engagement1 sourceThu, Sep 17, 2026, 06:46 AM
AI Overview
- China Telecom AI's Xing4.0-29B-A4B (formerly TeleChat) uses a 29B-parameter MoE with 4B active per token, 64 routed experts, and MLA attention.
- The model natively supports 256K context, extensible to 512K, and is the first model of this scale trained entirely on Ascend NPU with MindSpore.
- It scores 75.0 on SWE-bench Verified and 57.5 on Terminal-Bench 2.1, beating Gemma4-26B-A4B on both while trailing Qwen3.6-35B-A3B on SWE-bench.
- Ascend co-optimization — MoE communication tuning, selective recomputation, and fused mHC operators — lifted training throughput roughly 96% over out-of-the-box performance.
- It supports fine-tuning via LLaMA-Factory and MindFormers, inference via SGLang, vLLM, and KTransformers, and agent frameworks including OpenCode, Claude Code, OpenClaw, and Hermes.
Related
Releases·Model Release
Huawei open-sources openPangu-2.0-Pro, a 505B-parameter MoE model
Huawei has open-sourced openPangu-2.0-Pro, a Mixture-of-Experts model trained on Ascend hardware. With 505B total parameters and 18B activated, it supports a 512k context and was pretrained on 34T tokens. The release includes a tech report detailing its post-training methods.
Huawei has open-sourced openPangu-2.0-Pro, a Mixture-of-Experts model trained on Ascend hardware. With 505B total parameters and 18B activated, it supports a 512k context and was pretrained on 34T tokens. The release includes a tech report detailing its post-training methods.
89 engagement1 sourceredditFri, Jul 31, 2026, 06:47 AM
Releases·Model Releasefeatured
Sber's ai-sage releases GigaChat-3.5-Reasoning: 432B-A28B MoE under MIT license
AI-sage released GigaChat-3.5 Reasoning, a 432B-A28B MoE with Gated DeltaNet for long-context efficiency. Domain experts (code, math, general) were trained with CISPO and merged into one model via on-policy distillation. The team reports performance near DeepSeek V4 Flash Preview while emitting 37% fewer reasoning tokens. Weights are public on Hugging Face under MIT, with a hosted demo at giga.chat.
AI-sage released GigaChat-3.5 Reasoning, a 432B-A28B MoE with Gated DeltaNet for long-context efficiency. Domain experts (code, math, general) were trained with CISPO and merged into one model via on-policy distillation. The team reports performance near DeepSeek V4 Flash Preview while emitting 37% fewer reasoning tokens. Weights are public on Hugging Face under MIT, with a hosted demo at giga.chat.
217 engagement1 sourceredditThu, Sep 10, 2026, 12:19 PM
Releases·Model Release
Tencent releases ContextPilot models for proactive agent context management
Tencent released ContextPilot, a proactive context-management framework for long-horizon language-model agents, as open-weight checkpoints in 14B, 8B, and E4B sizes. The 14B model is a Qwen3-14B checkpoint trained to plan, maintain long-term memory, and offload less useful context while agents keep reasoning and using tools. Weights are on Hugging Face with an accompanying paper and code repository, giving practitioners a drop-in approach to context management in agentic systems.
Tencent released ContextPilot, a proactive context-management framework for long-horizon language-model agents, as open-weight checkpoints in 14B, 8B, and E4B sizes. The 14B model is a Qwen3-14B checkpoint trained to plan, maintain long-term memory, and offload less useful context while agents keep reasoning and using tools. Weights are on Hugging Face with an accompanying paper and code repository, giving practitioners a drop-in approach to context management in agentic systems.
6 engagement1 sourceredditMon, Aug 31, 2026, 08:56 AM
Releases·Model Release
Tencent Hy Team releases Hy3, a 295B-parameter MoE model
Tencent's Hy Team released Hy3, a 295-billion-parameter Mixture-of-Experts (MoE) language model with 21B active parameters and a 3.8B MTP layer. The model outperforms similar-sized models and rivals open-source models with 2-5x parameters. It is available on Hugging Face in full (598GB) and FP8 quantized (300GB) versions.
Tencent's Hy Team released Hy3, a 295-billion-parameter Mixture-of-Experts (MoE) language model with 21B active parameters and a 3.8B MTP layer. The model outperforms similar-sized models and rivals open-source models with 2-5x parameters. It is available on Hugging Face in full (598GB) and FP8 quantized (300GB) versions.
0 engagement1 sourcerssMon, Jul 6, 2026, 11:57 PM
Releases·Model Releasefeatured
Meituan LongCat 2.0 trains and runs on Chinese ASICs, challenging NVIDIA
Meituan's LongCat 2.0 flagship model was fully trained and inferenced on Chinese ASICs, marking a significant architectural and hardware shift. The release signals a potential erosion of NVIDIA's dominance in AI compute, as domestic alternatives prove viable for large-scale workloads.
Meituan's LongCat 2.0 flagship model was fully trained and inferenced on Chinese ASICs, marking a significant architectural and hardware shift. The release signals a potential erosion of NVIDIA's dominance in AI compute, as domestic alternatives prove viable for large-scale workloads.
1.5k engagement1 sourceyoutubeSat, Aug 8, 2026, 09:40 PM