XingChen-AGI Xing4.0-29B-A4B
One-line: China Telecom’s agent-oriented 29B-A4B MoE (TeleChat lineage) with published SWE-bench Verified 75.0 and τ-Bench 64.6 — first model of its scale trained natively on Ascend NPU. Lab-reported numbers; independent verification pending.
Coding benchmarks
| Benchmark | Score | Source |
|---|---|---|
| HumanEval | — | — |
| HumanEval+ | — | — |
| MBPP | — | — |
| LiveCodeBench | — | — |
| Aider polyglot | — | — |
| SWE-bench Verified | 75.0 | lab-reported (SWE-agent harness, 210K ctx) |
| SWE-bench Multilingual | 66.0 | lab-reported (SWE-agent harness, 210K ctx) |
Agentic / tool-use benchmarks
| Benchmark | Score | Source |
|---|---|---|
| BFCL | — | — |
| τ-Bench | 64.6 | lab-reported (τ³-Bench, pass^1 @ 4 runs) |
| ToolACE | — | — |
| GAIA | — | — |
| Terminal-Bench 2.1 | 57.5 | lab-reported (terminus-2, avg of 3 runs) |
| Claw-Eval | 76.6 | lab-reported (official harness, avg of 3 runs) |
| DeepresearchBII | 60.8 | lab-reported (OpenCode + Exa MCP) |
Reasoning (context only)
| Benchmark | Score | Source |
|---|---|---|
| MMLU / MMLU-Pro | — | — |
| GPQA Diamond | — | — |
| MATH | — | — |
| AIME 2026 | 90.0 | lab-reported (avg of 5 runs) |
| IFBench | 69.7 | lab-reported |
Efficiency
- Active params: 4B (29B total, 64 routed experts / 4 active + 1 shared, MLA attention, mHC + MTP)
- Context length: 256K (YARN-extended from 4096 pretrain)
- VRAM (fp16 rough): ~58 GB (needs quantization for single consumer GPU)
What makes it notable
First model of this scale trained end-to-end on Ascend 910C NPUs (MindSpore), from China Telecom’s Xing/TeleChat line. Card claims the strongest agentic numbers in its size class — SWE-bench Verified 75.0, Terminal-Bench 2.1 57.5, τ-Bench 64.6 — edging Qwen3.6-35B-A3B on SWE-bench Multilingual and beating Gemma4-26B-A4B across most agentic rows. Positioning is explicitly agent-loop: compatibility targets include OpenCode, Claude Code, OpenClaw and Hermes, with OpenAI-compatible API and vLLM/SGLang/KTransformers support.
Caveats: all numbers are lab-reported (no independent Artificial Analysis / leaderboard confirmation found yet); model_type: xing4_0 is a custom architecture (transformers 5.14.1, auto_map in-repo) — llama.cpp support unconfirmed, so local runs currently need vLLM/SGLang or community GGUF builds. Alibaba and Zhipu comparisons on the card are direct lab-harness runs, not standardized third-party results.