LLM Release Radar

New Open-Source Models

• 🔥 Chroma Context-1: Chroma introduced Context-1, a 20B parameter agentic search model with retrieval performance...

• 20B agentic search model achieves frontier-scale LLM retrieval at fraction of cost and 10x faster inference.
• HF weights live:...

TurboQuant breakthrough for local AI:

• Run same models with 4-8x larger context windows on consumer devices, using same or less memory.
• No accuracy...

Google launched Gemini 3.1 Flash Live, designed for real-time voice and vision agents with ultra-low latency and more natural conversations.

Key...

New Open-Source Model Releases

• 🔥 Minimax M 2.7: Minimax released M 2.7, an open-source self-evolving agent model with Artificial Analysis...

Japan's pure full-scratch LLM breakthrough:

• Release: PFN's PLaMo 3.0 Prime β dropped 2026-3-19, zero foreign models, strong in Japanese...

Google Research's TurboQuant slashes LLM KV cache memory by at least 6x with zero accuracy loss.

Practical wins for model testers:

• Extend...

Qwen3.5 leads open models for practical local agents:

• 1M+ token context handling in variants, plus native agentic abilities
• Reasoning/coding...

IBM and partners have contributed the llm-d project to CNCF, boosting open-source AI infrastructure for scalable LLM ops.

• Open-source agent model first to deeply self-evolve, excels in complex harnesses, agent teams, skills, dynamic tools
• Benchmarks: 50 on Artificial...

MiniCPM delivers giant-level performance on everyday hardware. Key practical details for local runs:

• Benchmarks: Rivals LLaMA and Mistral via...

New Open-Source Releases

• 🔥 Ai2 MolmoWeb: Allen Institute for AI released open-source MolmoWeb, a visual AI agent for web browser control built...

Ai2 launched MolmoWeb, an open-source visual AI agent on Molmo 2 (4B/8B params) for browser control—free weights, data, code (soon) for local/cloud...

Chroma study: 18 models show bigger context windows degrade performance every time.

WISC strategies boost agents:

• W-Write: Externalize memory.
  -...

• 560B open-source MoE bridges natural language to formal theorem-proving in Lean4
• Agentic tool-integrated Native Formal Reasoning decomposes into...

Open-Source Research Agents

• 🔥 MiroThinker-1.7 & H1: MiroThinker-1.7 improves step-level reliability through agentic mid-training with...

Rising actionable guides simplify local open-source LLMs:

• Hardware scanner: llm-checker CLI assesses your PC for Ollama models.
• Model installs:...

• MiroThinker-1.7 boosts step-level reliability via agentic mid-training for structured planning and tools
• H1 adds local/global verification to...

Yandex's Alice AI LLM, Russia's next-gen proprietary model, completed a full training cycle on its own data and infrastructure—initialized from custom base. Key peek into Russia's advanced, self-reliant AI stack.

Model Comparisons

• 🔥 GPT-5.4 mini vs DeepSeek V3.2: Comparison between GPT-5.4 mini (medium) and DeepSeek V3.2 (Reasoning) across intelligence,...