SmolLM3 3B
PulseAugur coverage of SmolLM3 3B — every cluster mentioning SmolLM3 3B across labs, papers, and developer communities, ranked by signal.
- 2026-05-12 research_milestone SmolLM3 3B achieved a top score in the Works With Agents benchmark for agent coding. source
4 day(s) with sentiment data
-
DoubleFree method speeds up LLM training by reusing exploration data
Researchers have developed a new method called DoubleFree to improve the efficiency of training multi-task Large Language Models (LLMs). This technique reuses historical trajectories from domain-specific reinforcement l…
-
Blog evaluates modern LLMs for poetry generation beyond GPT-2
A blog post details the process of selecting a modern Large Language Model (LLM) for a poetry generation project, moving beyond earlier models like GPT-2. The author evaluated six open-weight LLMs based on instruction f…
-
Open-source LLMs evaluated for Kinyarwanda-to-English translation
A developer explored the effectiveness of open-source language models for translating Kinyarwanda to English. Three models were evaluated: Qwen3-4B-Instruct-2507 from Alibaba Cloud, Llama 3.2 3B Instruct from Meta, and …
-
LLM researchers compare open-weight models for translation and recipe generation
Three research blogs explore open-weight LLMs for specific NLP projects, comparing models like Qwen2.5-1.5B-Instruct, Llama 3.2 3B Instruct, Gemma 3 4B Instruct, Qwen3-4B-Instruct-2507, Phi-4 Mini Instruct, and SmolLM2/…
-
Author compares Qwen3, SmolLM3, and Phi-4 for poem generation
The author is exploring three open-weight LLMs for a poem generation project: Qwen3-4B-Instruct-2507, SmolLM3-3B, and Phi-4-mini-instruct. Each model is evaluated based on its parameters, context window, licensing, and …
-
New OBU Framework Removes Hazardous LLM Capabilities Without Collateral Damage
Researchers have developed a new framework called Orthogonal Boundary Unlearning (OBU) to remove hazardous capabilities from large language models without harming their benign scientific utility. OBU works by identifyin…
-
Small open-weight LLMs benchmarked on Google Cloud TPU v5e
A guide details how to benchmark small, open-weight language models on a single Google Cloud TPU v5e chip using vLLM. The experiment uses Gemma 4 E2B as a baseline, evaluating models like Qwen3-4B, SmolLM3-3B, and Grani…
-
New RL Algorithm BoT-GRPO Enhances LLM Reasoning Capabilities
Researchers have introduced BoT-GRPO, a novel reinforcement learning algorithm designed to enhance reasoning capabilities in large language models. This method, an extension of GRPO, efficiently uses token-level rewards…
-
AI agents design their own inference hardware with openTPU project
An open-source project called openTPU demonstrates AI's capability to design its own inference hardware. The project, detailed in a single monorepo, includes hardware design in SystemVerilog, a simulator, and host softw…
-
Small language models learn to communicate and collaborate in TalkMesh system
Researchers have developed TalkMesh, a system that uses reinforcement learning to enable small language models to communicate and collaborate. This decentralized mesh of agents learns when and what information to share,…
-
New methods refine on-policy distillation for enhanced AI capabilities · 7 sources tracked
Researchers are exploring advanced techniques in on-policy distillation (OPD) to enhance language model capabilities by combining multiple "teacher" models. New methods like LEGO-OPD and SAKI focus on factorizing and ro…
-
New dataset AMPLE-Math probes value of privileged info in LLM self-distillation
Researchers have developed AMPLE-Math, a new dataset comprising over 5,000 mathematical problems, to investigate the impact of privileged information in on-policy self-distillation (OPSD) for language models. Their find…
-
Instella-MoE: New open-source MoE language model released
A new technical report introduces Instella-MoE, an open-source Mixture-of-Experts (MoE) language model with 16 billion total parameters. Trained on AMD Instinct GPUs, the model incorporates innovations like Gated Multi-…
-
New recipe trains AI models on consumer GPUs for under $7,000
Researchers have developed a cost-efficient pretraining recipe for language models, enabling training on consumer-grade hardware like RTX 5090 GPUs for under $7,000. This new method, demonstrated with the Puro-2B model …
-
Open-source recipe trains 2B LLMs on consumer GPUs for under $7K
Researchers have developed an open-source pretraining recipe that significantly reduces the cost of training large language models, making them accessible on consumer GPUs for under $7,000. Their Puro-2B model, trained …
-
webAI releases TwIL-LM formal logic models for local autoformalization
webAI has launched TwIL-LM, a family of two formal logic models available in 1.7B and 3B parameter sizes. These models are designed for autoformalization, translating English into first-order logic and verifying conclus…
-
New research tackles LLM alignment, safety, and optimization challenges
Researchers are exploring new methods to improve the alignment and reliability of large language models (LLMs). One study identifies a vulnerability in byte-pair encoding (BPE) tokenization that can be exploited to bypa…
-
Quantization and temperature effects on LLM safety analyzed
A new study investigated the combined effects of model quantization and sampling temperature on the safety alignment of large language models. Researchers found that standard quantization methods like INT4 and INT8 gene…
-
New Sparse Autoencoder Model Enhances LLM Feature Interpretability
Researchers have introduced the Sign-Aware Gated Sparse Autoencoder (SA-GSAE), a novel architecture designed to improve the interpretability of features extracted from Large Language Models. Unlike standard SAEs that en…
-
Tiny models outperform frontier AI in agent coding benchmark
A recent agent coding benchmark revealed that smaller, more efficient models are outperforming larger, frontier models. The SmolLM3 3B model, capable of running on a laptop, achieved a score of 93.3, significantly surpa…