Ling-3.0-tiny
One-line: Ant Group’s (InclusionAI) small-class hybrid-linear reasoning MoE — 7.9B total / 1.3B active (KDA + MLA, 3:1, 128 experts) — built explicitly for local/edge agent loops; AA Intelligence Index 25 (v4.1.1) / Agentic Index 16, ~168 tok/s, FP8 runs on a single DGX Spark or M4 Pro MacBook.
Coding benchmarks
| Benchmark | Score | Source |
|---|---|---|
| HumanEval | — | not published |
| HumanEval+ | — | not published |
| MBPP | — | not published |
| LiveCodeBench | — | not published |
| Aider polyglot | — | not published |
| SWE-bench Verified | — | not published |
| SWE-bench Multilingual | — | not published |
| Terminal-Bench 2.1 | — | evaluated under AA protocol (Terminus 2 harness, 256K ctx) — number only in image-only chart |
The card’s benchmark table is a single image (no .eval_results/ YAMLs shipped for tiny, unlike flash). Independent AA verification exists only as composite indices (below), not per-benchmark splits in extractable form. No HumanEval/MBPP/LiveCodeBench/Aider/SWE-bench numbers published anywhere — cells left — rather than invented.
Agentic / tool-use benchmarks
| Benchmark | Score | Source |
|---|---|---|
| BFCL | — | not published |
| τ-Bench | — | not published (AA Agentic Index includes τ³-Banking component, composite only) |
| ToolACE | — | not published |
| GAIA | — | not published |
| AA Agentic Index | 16 | Artificial Analysis (independent, composite) |
Reasoning (context only)
| Benchmark | Score | Source |
|---|---|---|
| AA Intelligence Index v4.1.1 | 25 | Artificial Analysis (independent; composite of GDPval-AA v2, τ³-Banking, Terminal-Bench 2.1, SciCode, HLE, GPQA Diamond, CritPt, AA-Omniscience, AA-LCR) — #6/56 in class, class median 8 |
| MMLU / MMLU-Pro | — | not published |
| GPQA Diamond | — | in image-only chart |
| MATH | — | not published |
| AIME 2025/2026 | — | in image-only chart |
| HLE | — | in image-only chart |
Efficiency
- Active params: 1.3B per token (8 of 128 routed experts + 1 shared expert; 3:1 KDA–MLA hybrid-linear, 24 layers)
- Total params: 7.9B (disk-resident footprint)
- Context length: 128K native (max_position_embeddings 131072), served at 262,144 via YaRN (rope_theta 6,000,000)
- VRAM (fp16 rough): ~16 GB fp16; FP8 ≈ 8 GB; card claims ~8.34 GiB peak memory at 8K ctx. Card-verified deployments: NVIDIA DGX Spark 100–105 tok/s (FP8), Apple Silicon M4 Pro MacBook 86–90 tok/s. BF16/FP8/INT4 weights all provided — genuinely single-box/consumer class.
- ⚠ custom architecture (
BailingMoeV3ForCausalLM,trust_remote_coderequired — custom model code in repo, not stock transformers). Ollama support in-flight (PR #17643, MLX on Apple Silicon only); SGLang has a dedicated cookbook + pre-built image; vLLM viavllm-ling-v3branch.
What makes it notable
A rare genuinely-small agentic MoE from a major lab: 1.3B active params puts it in the class of models that actually run on a laptop or edge box while scoring AA Intelligence Index 25 — well above its class median (8) and in the same class as much larger reasoning models, per independent AA testing. The hybrid KDA+MLA architecture (borrowed from Kimi Delta Attention + MLA) and 128-expert sparse MoE are the same family as the much larger Ling-3.0-flash (124B), shrunk to consumer hardware — flash’s eval files proved the family’s reasoning profile (SWE-bench Pro 56.6, AIME 93.2); tiny ships no machine-readable evals, so treat agentic claims as lab-positioned with only AA’s composite as independent verification. Positioning vs local leaders: for a 1.3B-active footprint this is one of the strongest agentic-on-paper options (free on OpenRouter until 8/14), but the missing per-benchmark numbers and the custom-code requirement mean verify it yourself before building agent loops on it. Native thinking mode + ling3 tool-call parser (SGLang/vLLM) make it directly agent-loop deployable.