ik_llama.cpp PR #2808 adds native runtime for China Telecom's Xing4.0-29B-A4B MoE with mHC and MLA
Any_Bus9205 · reddit · 2026-10-31
China Telecom's TeleAI released Xing4.0-29B-A4B, a 29.5B MoE (4B active) combining DeepSeek-style MLA attention with 4-stream mHC residual mixing. With upstream llama.cpp PRs stalled over debates on model-specific GGML opcodes, the author opened PR #2808 on ikllama.cpp for full native support:
- Full runtime: 4-stream mHC built entirely from generic GGML primitives (affine transforms, clamp, Sinkhorn iterations) — zero new ops, runs on CPU/CUDA/Metal/Vulkan out of the box; MLA attention with dynamic head sizing; dense lead layers + routed + shared experts
- Converter: HF-to-GGUF script with native SentencePiece extraction and direct mHC tensor mapping
- Quantization: protects hcattnfn/hcffnfn from quantization; keeps MLA latent projections at source precision for Q6K/Q80
- Verification: perplexity 12.4190, coherent generation, full chat template with thinking tags and tool calling
Full build/convert/run commands included.
More from Models
- Gemini 4 Carbon takes #1 on Vending-Bench 2 at lower cost than rivals — vedantmisra · 2026-10-31
- Particle-1, a token-efficient model, launches on Cloudflare AI Gateway — garrytan · 2026-10-31
- OpenAI Claims Proof That L = BPL, Hinting at AI-Generated Formal Mathematics — chaumian · 2026-10-31
- Users report AI models increasingly 'assuming' facts and swapping queries — abhishek_8899 · 2026-10-31
- Matt Wolfe weekly: Anthropic IPO filing, GPT-6.1 Terra and Google Genie 4 lead AI news — Matt Wolfe · 2026-10-31
- Qwenputer: a full RISC-V CPU inside Qwen3.5-4B weights, running C in forward passes — armand_ruiz · 2026-10-31