#open-source-llm

Kimi K2.7-Code is Moonshot's new open-source coding model: +21.8% on Kimi Code Bench v2, 30% fewer reasoning tokens, plus Kimi Code CLI plans from $19/mo.

DeepSeek-V4 ships April 24, 2026 as open-weight MoE: Pro (1.6T/49B active) and Flash (284B/13B), 1M context, 27% FLOPs and 10% KV cache vs V3.2.

Moonshot's Kimi K2.6 ships 300-agent swarms, 12-hour coding runs, WebGL hero sections, and open-source SOTA on SWE-Bench Pro. Agency playbook and benchmarks.

Qwen 3.5 medium series: Flash, 35B-A3B, 122B-A10B, and 27B. Benchmarks vs GPT-5 mini and Claude Sonnet 4.5, pricing from $0.10/M tokens.

Qwen 3.5-397B scores 83.6 on LiveCodeBench v6 and 91.3 on AIME26 with 17B active MoE params. Benchmarks vs GPT-5.2, Claude, and pricing details.

DeepSeek V4 brings 1 trillion parameters, 1M token context, and Engram O(1) memory. Architecture details, leaked benchmarks, and what it means for developers.

DeepSeek R1 revolutionized open-source AI with reasoning capabilities. Compare it to Qwen 3's 1T parameters and Mistral Large for enterprise deployment.

MiniMax M2.1 is a real Dec 2025 coding and agentic workflow model listed on OpenRouter, with 10B active parameters and Digital Employee positioning.

Deploy Llama 3.3, Mistral 3, Qwen 3 locally with Ollama, LM Studio, or vLLM. Hardware requirements, quantization, and enterprise self-hosting patterns.

GLM-4.7 achieves 73.8% SWE-bench and 87.4% τ²-Bench with Preserved Thinking. Complete developer guide for the $3/month Claude Code alternative.