Xingchen unveils Xing 4, a 285B-parameter model built on Ascend chips
Xingchen AGI Lab has launched Xing 4, a 285-billion-parameter language model family trained entirely on Ascend 910C chips. The TeleChat successor uses MindSpore HyperParallel across 8,192 cards and adds TeleGym for agent post-training. A 29-billion-parameter MoE member is being open-sourced.
- Original source: blockchain.news.
- The source published it on Nov 18, 2026 at 6:36 AM.
- No second outlet has reported it yet. Treat it as a single-source claim.
Xingchen AGI Lab pulled the wraps off Xing 4, a family of large language models that pushes the TeleChat line into 285-billion-parameter territory. The research landed on arXiv on November 17, 2026, under paper number 2611.22115, and it is aimed squarely at domestic silicon — Ascend 910C chips rather than imported accelerators.
The training stack is built on MindSpore HyperParallel, a framework the lab tuned for large-scale distributed work. It leans on topology-aware hybrid sharding, efficient distributed optimizer execution, and CPU affinity with core isolation to keep scheduling interference down. Those tweaks let Xingchen run pretraining across 8,192 Ascend 910C cards, handling long-sequence workloads and activation memory management along the way.
On architecture, the team balanced sparse activation, attention design and speculative decoding to squeeze out compute and memory efficiency. A mixture-of-experts topology paired with lightweight correction heads sharpens speculative decoding, which the researchers say keeps execution stable in long-context scenarios while supporting reasoning and agentic tasks.
Post-training runs through TeleGym, a unified agent execution platform that bundles execution, evaluation and trajectory construction with policy optimization. Disaggregated asynchronous execution, bounded oversampling and partial rollout let it cope with tasks of wildly different durations without breaking trajectory continuity across policy updates.
Xingchen describes the work as a step toward building advanced AI on homegrown hardware, giving future LLM development a scalable base for complex reasoning and multi-step planning. The lab is also open-sourcing a 29-billion-parameter member of the mixture-of-experts family.
TopicsXingchen AGI Lab · Xing 4 · TeleChat · Ascend 910C · MindSpore HyperParallel · TeleGym
Source: blockchain.news
Written by AI from verified reporting.
More stories
- Tech & AI coindesk.com Caught by @gimbung 0 viewsGoogle finds an AI that hunts and exploits software flaws without a human
- Tech & AI x.com 1 viewCalifornia AG says his office is looking into the Nvidia scandal
- Tech & AI cryptobriefing.com 1 viewNvidia's Singapore Unit Books $104B for 2025 as Reporting Rules Shift
- Tech & AI x.com 1 viewNvidia Board Faction Pushes for Jensen Huang's Return as CEO, per Semafor

