AI Briefing#open-source-llm
The latest AI and developer news about #open-source-llm, with the original source and a short summary.
The latest AI and developer news about #open-source-llm, with the original source and a short summary.
SFT on 52.8 million tokens reduced the Chinese government narrative adoption rate from 22% to 0.9%.
The model achieves 80+ tokens/s decoding and 29GB peak memory usage on compatible laptops, outperforming several cloud models on coding and reasoning benchmarks.
Egzakta releases Kvark Sparkle 1.1, an Apache 2.0 fine-tune of Qwen3.8-27B that correctly declines answers not found in documents 91.5% of the time.
The model reduces propaganda adoption from 22% to 1% and improves fact scores from 2.32 to 2.98 while maintaining agentic capabilities.
A fine-tuned 20B model using a designer-coder split and check-repair harness achieves a 59.7 mean score on PresentBench, ranking second overall and first in Economics and Education.
The open-weight model achieves a 13.7s median response time versus 28.1s for the base model while maintaining comparable accuracy on five benchmarks.
The new model retains the Spark-X2.5-4B architecture while delivering state-of-the-art performance among similarly sized open-source models.
The new version eliminates repetition loops found in the preview, achieves 0/50 loops on SWE-bench tasks, and introduces an open-source engine enabling 262K context within a 16 GB memory budget.
The Apache-2.0 licensed model uses a hybrid architecture with 54 KDA layers and 17 BCSA layers to minimize KV cache usage while supporting a 262K context window.
AlephAlpha has released Kolibri, an open-weight Mixture-of-Experts model with 78.1B total parameters and 3.46B active parameters, designed for sovereign and regulated domains with a focus on German language performance.
By modifying only 0.02% of the parameters of an open model, it achieved first place in five categories, including a perfect score on AIME2026, and released the weights.
The repository lists 27 models, including specialized fine-tunes like DeepHat V2 and GLM-5.3 variants, alongside deployment guides for local and cloud inference.
Next GenAi has released 'Rebekah 1.8B', an open-source LLM for edge deployment focused on data curation.
According to a Mozilla report, the performance gap between open and closed models has narrowed to an average of about 4 months, with competition shifting from the models themselves to harnesses.
Empirical data demonstrates that on OpenRouter, even for the same model, performance and features such as benchmark scores, vision recognition, and parameter support can vary significantly depending on the Provider.
IFM has open-sourced K2 Horizon, a suite of 6 models that exposes the entire process from pre-training to agent post-processing.
Ant Group has released Ling-3.0-flash-Fin, a 124B-parameter open-source model trained on financial data.
The Apodex team released Apodex 1.1, an open-source model specialized for agent intelligence, along with its harness.
Ornith-1.5 achieves open-source SOTA through self-improvement via self-generated tasks and reinforcement learning
An open-source 30B model surpassed GPT-5.5 on the VibeWorlding benchmark, setting a new standard for 3D world generation.
LG AI Research has released its open-source model K-EXAONE 2.0 via Hugging Face.
Through low-cost RL fine-tuning, a 9B-scale open model outperformed high-performance models on specific domain performance.
Soofi S, a highly efficient hybrid Mamba-MoE open-source model specialized for German and English, has been released.
LG AI Research has released 3 EXAONE 3.5 open-source models that maximize performance and efficiency.