Tag
作者探讨了一种面向本地受限硬件（<6GB VRAM）的解耦式多智能体架构：用单个中等规模推理模型作为编排者，配合多个1-3M参数的CPU端「纳米模型」处理高频例行任务，仅在低置信度时向上升级，实现零云、空气隔离的边缘优先系统。
作者训练了一个 414k 参数的从零实现的 Transformer 来模仿 boids 鸟群行为，并用线性探针研究模型内部表征的三条 boids 规则是否与其实际使用机制一致，发现 cohesion 可在训练前读出、alignment 依赖第二注意力层，且首次训练的模型实际并未利用鸟群信息。
Cactus Compute releases Whistle: a 16.9 MB speech recognition model that runs on pure CPU, supporting seven European languages. Measured first-token latency of 5-8 ms and decoding speed of ~1300 tokens/sec. However, for unsupported languages (such as Chinese and Japanese), it does not throw errors — instead it silently outputs fake words, posing a risk of silent data corruption in production environments.
这篇论文在极小规模（8.7M 参数、4 层）的中文 BERT 上严格控制变量，对比了 MLM、WWM 和 MacBERT 三种预训练策略，发现小规模下的策略排名（MLM > WWM ≫ MacBERT）与大规模结论相反，并指出 MacBERT 在同义词词典覆盖不足时会出现困惑度严重退化、训练损失不可靠的问题。
The author questions why tiny AI models with fewer than 50M parameters or swarms of specialized micro-models are rarely deployed in production, speculating on reasons like tooling biases or the convenience of generalist models.
A tweet from @tobi discusses how training tiny models for specialized use cases with a self-improving flywheel is highly effective, noting that Shopify ML team's finetuned 0.8b model outperforms GPT 5.6-sol xhigh in a specific task.
TwiL-LM, a family of tiny specialized open models, is released on HuggingFace. The 3B version outperforms OpenAI's 120B gpt-oss on formal reasoning benchmarks and runs efficiently on consumer hardware.
A developer demonstrates running a 28.9 million parameter language model on an $8 ESP32-S3 microcontroller using Google's Per-Layer Embeddings to store most parameters in flash, achieving around 9.5 tokens per second on-device text generation.
SupraLabs releases reasoning-corpus-4K-5M-v1, a 5M-sample reasoning dataset for training small language models (SLMs), featuring chain-of-thought traces and ChatML format, hosted on Hugging Face.
TinyRouter is a tiny 10K-parameter LLM router that learns to route each question to the best specialist model from a pool of open-source LLMs, using evolutionary training. It achieves performance matching or exceeding individual models on MMLU and math benchmarks.
A student introduces Silia, a novel transformer architecture that combines attention and FFN into a unified operation to save parameters at scales ≤10M, achieving comparable performance to GPT-2 with fewer parameters despite limited compute resources.
Announces an upcoming video on training tiny models for preference tuning, covering reward models, RLHF, DPO, ORPO with Unsloth and TRL.
This paper presents Flavors of Moonshine, a suite of tiny specialized ASR models for edge devices. The authors show that monolingual models trained on a balanced mix of human-labeled, pseudo-labeled, and synthetic data outperform larger multilingual models like Whisper, achieving state-of-the-art error rates for small models and enabling on-device ASR for underrepresented languages.