The big releases of June 2026 are all frontier-tier: GLM-5.2, Kimi K2.7 Code, minimax M2.7, and Qwen3-Coder 480B. Plus a few smaller wins for the rest of us. Here’s what shipped, what matters for local deployment, and how to try each.
Frontier tier (256GB+ VRAM or multi-GPU)
GLM-5.2 from Z.ai
Released: June 16, 2026
Architecture: 744B parameters, 40B active MoE
License: MIT
Context: 128K
Strengths: Z.ai’s most-reasoning-capable model. State-of-the-art on SWE-Bench Pro. Designed for long-horizon agentic engineering and complex reasoning.
Run it locally: ollama pull glm-5:744b (Q4 needs ~410GB VRAM, FP8 needs ~830GB)
Watch for: GLM-4.7 Flash (30B), released the same week, fits on a single 24GB card and is the strongest 30B coder in mid-2026.
Kimi K2.7 Code from Moonshot AI
Released: June 12, 2026
Architecture: MoE (full size not yet disclosed)
License: Modified MIT
Context: 128K
Strengths: Code-specialized agent model. Strong on multi-file refactors and long-horizon coding workflows. Updated from Kimi K2.6.
Run it locally: mullama pull kimi-k2-7-code (full size needs multi-GPU; expect distilled checkpoints soon)
Watch for: The Kimi K2.7 base model (non-Code) is expected to follow.
minimax M2.7
Released: March 18, 2026 (rolling into June as the de facto standard)
Architecture: 480B parameters
License: minimax License (open-weight)
Context: 128K
Strengths: Among the most-pulled models in the Ollama library in 2026. Exceptional multilingual capabilities. Designed for coding, agentic workflows, and professional productivity.
Run it locally: ollama pull minimax-m2-7 (480B at Q4 needs ~280GB)
Watch for: minimax M3 was released same day with similar specs but improved agentic workflows.
Qwen3-Coder 480B
Released: Late 2025, GA in 2026
Architecture: 480B MoE
License: Custom (research-friendly)
Context: 256K
Strengths: Code agent-focused. Long context. Strong on multi-step refactors and codebase-wide changes.
Run it locally: ollama pull qwen3-coder:480b (Q4 needs ~280GB)
Single-GPU tier (24GB VRAM)
GLM-4.7 Flash (30B)
Released: June 2026
Architecture: 30B dense
License: MIT
Context: 128K
Strengths: The strongest 30B class model in 2026. Strong on coding, agentic, and general chat. Fits in 24GB.
Run it locally: ollama pull glm-4-7-flash — fits in 18GB at Q4, leaves 6GB for context.
Qwen 3 32B (continued dominance)
The Qwen 3 32B and Qwen 3 30B-A3B MoE remain the go-to for 24GB GPUs. GLM-4.7 Flash is now a credible alternative; pick by use case (GLM for coding/agentic, Qwen for general).
Edge tier (phones, Raspberry Pi, Steam Deck)
LFM2.5 Thinking (1.2B)
Released: January 2026
Architecture: 1.2B hybrid (Liquid AI)
License: Apache 2.0
Context: 32K
Strengths: Adds reasoning to a tiny footprint. Best 1B-class model in 2026. Runs on phones and edge devices.
Run it locally: ollama pull lfm2-5-thinking — fits in 1GB of RAM, runs at 10-20 tok/s on a Steam Deck.
What it all means for local AI in 2026
The local AI ecosystem now has clear three-tier structure:
- Frontier tier: GLM-5.2, Kimi K2.7, minimax M2.7, Qwen 3 480B+ — frontier-class, needs 256GB+ VRAM or multi-GPU
- Single-GPU tier (24GB): Qwen 3 32B, GLM-4.7 Flash, DeepSeek R1 32B distilled — best in class for 3090/4090/5090
- Edge tier (phones, embedded): LFM2.5 Thinking 1.2B, Llama 3.2 1B/3B, Qwen 2.5 0.5B/1.5B — fits anywhere
For most local AI deployments in 2026, the single-GPU tier is the sweet spot. The 30B-class models are strong enough to handle most workloads, and 24GB cards are affordable.
See also
- Best local LLM 2026 — annual best-of roundup
- GLM-5.2 model page — full setup details
- Kimi K2.7 Code model page — code agent specs
- Local LLM on RTX 3090 — 24GB build recommendations
About this series
Local AI News is a monthly release roundup covering the most important local AI releases, with setup commands and VRAM requirements. Updated on the first Monday of each month. Subscribe via RSS.