GROUNDING

Open Source LLMs

6 min read

Open Source LLMs are language models released with open weights — and sometimes training code or data — so builders can inspect, fine-tune, self-host, and run them without depending on a hosted API.

The appeal is control: data privacy, predictable cost at scale, customization through fine-tuning, and no vendor lock-in. The tradeoff is that the operator now owns inference — the hardware, serving stack, and optimization that a closed API would handle.

For builders the decision hinges on whether control and privacy outweigh that operational burden; open models now rival closed ones on many tasks but trail at the frontier. Open weights also rarely means fully open, since licenses and undisclosed training data vary, so terms matter as much as benchmarks.

Topic: Models Related: Code Agents RAG

Recent Updates

  • 2026-07-14: Picchio benchmarks local LLMs with effective bits-per-weight, three speed lanes, and GPU fallback detection (Hacker News) · github.com — Hugging Face · LM Studio · NVIDIA · Qwen3.5-9B
  • 2026-07-15: AINews digest: Codex usage surge, agent harness observability, and heavily quantized open models for local agents (Latent.Space) · latent.space — OpenAI · JetBrains · LangChain · PrismML · Tencent OpenMOSS Locally AI · Latent.Space Richard MacManus Addy Osmani · GPT-5.6 · Qwen 3.6-27B · Bonsai 27B Hy3 · Gemma 4 · Qwen3.5-122B-A10B · GLM-4.7-Flash · GLM-5.2 · DeepSeek-V4-Flash · Mimo v2.5 MOSS-VL-Realtime
  • 2026-07-15: LLaMA 3 (8B) fine-tuned as an efficient RAG reranker to replace cross-encoders (cs.LG updates on arXiv.org) · arxiv.org — Shreeya Dasa Lakshminath LLaMA 3 (8B)
  • 2026-07-15: Function-Aware Fill-in-the-Middle Mid-Training for Coding Agent Foundation Models (cs.CL updates on arXiv.org) · arxiv.org — GitHub Qwen2.5-Coder-Instruct · Qwen3-8B
  • 2026-07-15: On-Device Deep Research at 4B: Exposure Bounds Faithfulness, Retrieval Bounds Coverage (cs.CL updates on arXiv.org) · arxiv.org — Vinay Kumar Chaganti
  • 2026-07-15: LLaMA 3 (8B) fine-tuned as an efficient RAG reranker replacing cross-encoders (cs.CL updates on arXiv.org) · arxiv.org — Shreeya Dasa Lakshminath LLaMA 3 (8B)
  • 2026-07-15: Thinking Machines Lab Releases Inkling, Competitive Open-Source Multimodal Model (Все статьи подряд / Искусственный интеллект / Хабр) · habr.com — Thinking Machines Lab · OpenAI · Anthropic · Google · Moonshot AI · Hugging Face · NVIDIA · Cognition Mira Murati · Inkling · kimi-k2.5 · kimi-k2.6 · GLM-5.2 · Nemotron 3 Ultra · DeepSeek-V3 · Claude Opus 4.6 · Gemini 3.5 Flash · Gemini-3.1-Pro · Grok 4.3
  • 2026-07-16: Auditing open-weight LLM behavior with persona vectors (cs.CL updates on arXiv.org) · arxiv.org
  • 2026-07-16: Thinking Machines Lab releases Inkling, an open-weights multimodal model family (Latent.Space) · latent.space — Thinking Machines Lab · Latent.Space · Hugging Face · vLLM · SGLang · Modal · Baseten · Databricks Mira Murati Soumith Chintala John Schulman Lilian Weng · Inkling · Inkling-Small
  • 2026-07-16: Hugging Face reports a July 2026 security incident and AI-assisted response (Hugging Face - Blog) · huggingface.co — Hugging Face · GLM-5.2
  • 2026-07-16: European AI is showing up more in enterprise assistants than consumer chat products (Все статьи подряд / Искусственный интеллект / Хабр) · habr.com — Mistral SAP · OpenAI · Google · Meta · Anthropic SAP Sapphire SAP Business AI Platform · Deutsche Telekom Business GPT Buhl WISO Steuer Taxy.io GmbH German Commons SpeakLeash Cyfronet AGH PLGrid · Hugging Face InPost · Sber Joule · Claude · Gemini Meta Llama Teuken-7B · Bielik Mistral 7B v0.2 · GigaChat
  • 2026-07-16: Omar Sar argues builders should combine models and rely on private evals (@dair_ai) · x.com — DAIR.AI · Fable · Mythos 5
  • 2026-07-17: Small language models are already running on phones and in browsers (Все статьи подряд / Искусственный интеллект / Хабр) · habr.com — DeepMind · Microsoft · Apple Georgiy Gerganov · Chinchilla Gopher · TinyLlama · Phi · LLaMA
  • 2026-07-18: The Kimi K3 Moment (Hacker News) · stephen.bochinski.dev — Anthropic · OpenAI Semgrep · Kimi K3 · Claude · Fable · GLM-5.2 · Opus · GPT-5.6
  • 2026-07-19: Last Week in AI: China, Compression, and the Open-Model Race (TheSequence) · thesequence.substack.com — TheSequence · Thinking Machines Lab · Moonshot AI · PrismML · OpenAI · NVIDIA · Google Shanghai World AI Conference Xi Jinping · Inkling · Kimi K3 · Bonsai 27B GPT-Red · GPT-5.1 · GPT-5.6
  • 2026-07-20: Hugging Face describes an AI-driven intrusion and its incident response (‌Z.ai) · huggingface.co — Hugging Face · z.ai ZixuanLi_ · GLM-5.2
  • 2026-07-20: Kimi K3 is presented as an open-weights frontier model (Interconnects AI) · interconnects.ai — Interconnects AI Moonshot AI · Anthropic · OpenAI · SpaceXAI · xAI Zhipu (Z.ai) · Meta · DeepMind · Alibaba Vals AI · Artificial Analysis Frontend Code Arena · Kimi K3 · DeepSeek R1 · Claude Fable 5 · GPT-5.6 Sol · Grok 4.5 · GLM-5.2 · Muse Spark 1.1 Gemini Flash 3.5 · Qwen 3.7 Max Qwen 3.8
  • 2026-07-20: Nativ: run open models locally on Apple Silicon Macs (Hacker News) · blaizzy.github.io — Nativ Google · Cohere Liquid AI · Apple MLX-VLM
  • 2026-07-21: Benchmarking and Fine-Tuning Sub-3B Open-Weight Models for Structured Local Tasks (cs.AI updates on arXiv.org) · arxiv.org — NVIDIA Qwen Coder 3B · Qwen2.5-1.5B · Qwen3.5-2B Granite 3.3 2B · SmolLM2-1.7B SmolLM2 360M SmolLM2-135M
  • 2026-07-22: Fine-tuned open-weight LLMs are tested for classifying vulnerability indicators in UK police logs (cs.CL updates on arXiv.org) · arxiv.org
  • 2026-07-22: Google expands Gemini Flash line while Poolside ships an open coding model (Все статьи подряд / Искусственный интеллект / Хабр) · habr.com — Google · Anthropic · Poolside · Hugging Face · NVIDIA · Thinking Machines Lab · Forbes · Artificial Analysis · Z ai Logan Kilpatrick William Alsup Andrea Bartz Jason Warner Elon Musk · Gemini 3.6 Flash · Gemini 3.5 Flash-Lite · Gemini 3.5 Flash Cyber · Gemini 3.5 Pro Gemini 4 · Grok 4.5 · GPT-5.6 Luna Meta Muse Spark · GLM-5.2 · Claude · Laguna S 2.1 · Inkling
  • 2026-07-23: A case study on using local LLMs for text-to-SQL in a WMS billing workflow (Искусственный интеллект – AI, ANN и иные формы искусственного разума) · habr.com — Аэросиб-С Ollama
  • 2026-07-25: Andrew Ng’s OpenWorker is an open-source desktop agent with local-first controls (量子位) · qbitai.com — GitHub · Slack · Jira · Notion · Linear · Hubspot Outlook · Gmail Google Calendar · OpenAI · DeepSeek · Ollama · Google · Baidu DeepLearning.AI Coursera · Stanford University Google Brain Andrew Ng Daphne Koller 文婷 · GPT-5.6 Sol · Claude Fable Gemini 3.6 · Kimi · GLM · Inkling
  • 2026-07-27: Kimi-K3: Open-Source Frontier Model with Native Agentic and Coding Capabilities (Hacker News) · huggingface.co — Moonshot AI · Hugging Face · Kimi K3
  • 2026-07-28: Moonshot AI releases Kimi K3 model weights with restrictive licensing (Simon Willison’s Weblog) · simonwillison.net — Moonshot AI · Hugging Face · OpenRouter · Kimi K3 · Kimi K2

FAQ

What is Open Source LLMs?

Open Source LLMs are language models released with weights, code, or permissive access that builders can inspect, fine-tune, or self-host. GROUNDING tracks open model releases, inference stacks, and benchmarks.

Which topic does Open Source LLMs belong to?

On the GROUNDING radar, Open Source LLMs is grouped under the Models topic.

Related concepts tracked by the radar include Code Agents, RAG.

Incoming · private radar

A radar tuned to your space.

Private, source-tuned weekly digests of the signals and knowledge gaps that matter to you — every claim human-reviewed before it ships.

Request a custom radar
$300–500/mo · 1–2 design partners

Graph View

Backlinks