As of July 2026. Four open-weight model families launched in June 2026, and they split cleanly into one model most people can actually run locally, and three that need serious hardware. Here’s the practical breakdown: what each model is, what it scores on the benchmarks that matter for local AI, how much memory it needs, and what hardware makes sense for it.
The one you can run: Gemma 4 12B
Google released Gemma 4 12B on June 3, 2026 under Apache 2.0 (per the official blog and model card). It’s an 11.95-billion-parameter dense model that processes text, images, audio, and video natively and runs on any machine with 16 GB of RAM or VRAM (reported, per the Ars Technica hands-on and VentureBeat coverage, June 2026). At roughly 8 to 9 GB in 4-bit quantization, it leaves headroom on a 16 GB machine for context, system overhead, and a browser.
If you have a 16 GB laptop and want a genuinely capable multimodal model that loads in standard stacks (llama.cpp, MLX, vLLM, Hugging Face Transformers) without a dedicated GPU, this is the June launch for you. For a full comparison against Qwen3.5-9B on the same hardware, see our best local model for a 16 GB laptop guide.
Gemma 4 12B is already in the model picker if you want to see how fast it runs on your exact machine.
The three you need a cluster for
MiniMax M3
MiniMax shipped M3 on June 1, 2026, a mixture-of-experts model with roughly 428B total parameters and about 23B active per token, built on MiniMax Sparse Attention for 1M-token context (reported, per the official Hugging Face repository and MarkTechPost, June 2026). It handles text, image, video, and computer-use tasks natively. Reported benchmarks: 59.0% SWE-Bench Pro, 66.0% Terminal Bench 2.1, 83.5% BrowseComp (reported, per MiniMax’s published results).
At a 4-bit quantization, M3 needs roughly 230 to 250 GB of memory, about 125 GB at 2-bit. That puts it in multi-GPU or high-memory capacity-box territory, not a laptop or single consumer card. If you are building a cluster and want to see whether a planned multi-GPU setup can clear that bar, the hardware compare view lets you stack memory across cards.
NVIDIA Nemotron 3 Ultra
NVIDIA launched Nemotron 3 Ultra on June 4, 2026 at Computex. It’s a 550B-parameter mixture-of-experts model with 55B active per token, using a hybrid Mamba-Transformer architecture (rather than a pure transformer) with native multi-token prediction and a 1M-token context (reported, per NVIDIA’s release materials and MarkTechPost, June 2026). The weights ship under the OpenMDW (Linux Foundation) license along with training data and recipes, an unusually open move for a frontier-scale release.
At roughly 300 to 340 GB for a 4-bit load, this is a datacenter model targeting 8-way H100 or B200 deployments. NVIDIA reports 300-plus output tokens per second on its own stack (reported, per the launch announcement, June 2026). It isn’t a local-workstation model; it competes with closed frontier APIs on the cloud side.
Kimi K2.7 Code
Moonshot AI released Kimi K2.7 Code on June 12, 2026 under a Modified MIT license. It’s a 1T-parameter MoE with 32B active (384 experts, top-8 plus one shared), a 256K context, and a small vision encoder for image and video input (reported, per Moonshot’s model card and Hugging Face repository, June 2026). The pitch is token efficiency: about 30% fewer reasoning tokens than K2.6 for the same task, and a jump from 50.9 to 62.0 on Kimi Code Bench v2 (reported, per the MarkTechPost coverage, June 2026). At 81.1 on MCP Mark Verified it edges past Claude Opus 4.8 in that benchmark (reported, per Moonshot’s published results).
At roughly 540 GB for a 4-bit load (about 250 GB at 2-bit), K2.7 Code is homelab-or-bigger. Independent third-party SWE-Bench and Terminal-Bench results weren’t yet available as of this writing.
The takeaway
June 2026 shows the local AI market splitting into two lanes more clearly than ever. On one side, a dense 12B model gets genuinely capable multimodal inference onto a standard laptop for the first time. On the other, the frontier keeps getting bigger and more efficient per active parameter, but those models stay in the 200 GB-plus tier where they need a memory-pooled build or a multi-GPU cluster.
If you are comparison shopping between capacity boxes or multi-GPU rigs and want to know which of these models fits your build, the model picker matches every open-weight model against your machine’s memory and shows the expected speed.
Sources
- Google Gemma 4 launch blog (blog.google, June 3, 2026)
- Gemma 4 model card (ai.google.dev, June 3, 2026)
- Ars Technica hands-on (arstechnica.com, June 2026)
- VentureBeat coverage (venturebeat.com, June 2026)
- MiniMax M3 Hugging Face repository (huggingface.co/MiniMaxAI/MiniMax-M3, June 1, 2026)
- MarkTechPost MiniMax M3 coverage (marktechpost.com, June 2026)
- MiniMax Sparse Attention paper (huggingface.co/papers/2606.13392)
- NVIDIA Nemotron 3 Ultra release materials (developer.nvidia.com, June 4, 2026)
- MarkTechPost Nemotron 3 Ultra coverage (marktechpost.com, June 2026)
- Kimi K2.7 Code model card (huggingface.co, June 12, 2026)
- MarkTechPost Kimi K2.7 coverage (marktechpost.com, June 2026)
- all benchmark scores marked “reported” per each vendor’s published results