Ling-3.0-tiny Released: 1.3B Active Params Beats GPT-OSS 120B
solyarisoftware · x · 2026-08-11
Ling-3.0-tiny is an open-weight model designed for reasoning, tool calling, and agentic workflows. It uses a 128-expert sparse MoE architecture with 7.9B total parameters but only 1.3B active parameters per token.
- Performance: Scores 73.4 on GPQA Diamond and 71.0 on IMO-AnswerBench. It reportedly beats OpenAI's GPT-OSS 120B (high) on Artificial Analysis despite its small size.
- Local Deployment: Available in BF16, FP8, and INT4. The INT4 version requires only 5.8GB memory, while FP8 peaks at 8.34 GiB for 8K context.
- Inference Speed: Achieves 100-105 tok/s on DGX Spark and 86-90 tok/s on M4 Pro MacBook.
- Runtime Support: Currently supports SGLang, vLLM, and MLX on Apple Silicon. Support for llama.cpp and LM Studio is pending.
More from Infra
- GitHub rewrote the 800k-line Copilot runtime in Rust using AI agents, cutting CPU and memory by over 90% — bibryam · 2026-10-16
- NVIDIA quietly delists the entry-level RTX 5050 from its marketplace, following the RTX 5090 — chemist_slime · 2026-10-16
- Grok claims 816B MoE model runs locally on dual M5 Ultra Mac Studios via MLX — grok · 2026-10-16
- Same 1,000 Cloudflare Durable Objects: $4,050/month cut to $22 with one trick — threepointone · 2026-10-16
- NVIDIA ships FP4-quantized Qwen3.8-Flash for local use, trending on Hugging Face — nvidia · 2026-10-16
- Baidu's Virtual Head Attention tames the KV cache, FLOPs, quality triangle in long-context LLMs — baidu · 2026-10-16