The past few weeks brought a wave of new open-weight LLMs. Unlike closed API-only models, open-weight models let you download the weights and run them on your own hardware or cloud – giving you control over cost, latency and data privacy. Here’s a quick developer’s guide to the most important releases and how to pick one.

The New Open-Weight Models at a Glance

Model	Developer	Architecture	Highlights
Mistral Large 4	Mistral AI	MoE, ~1T parameters	Flagship-scale open model
Muse Spark Open	Meta	MoE, 816B total / 42B active	Large MoE with modest active compute
Kimi 3.1 Pro	Moonshot AI	Open-weight	“Highspeed” variant is 2–3x faster
Flint	Recursive	MoE, 40B total / 2.5B active	Small active size, runs on modest hardware

All figures are as reported at launch.

Why Mixture-of-Experts (MoE) Dominates

Almost every new open model uses a Mixture-of-Experts architecture. Instead of activating every parameter for every token, an MoE model routes each token to a few “expert” sub-networks. That’s why Muse Spark Open has 816B parameters in total but only 42B active per token, and Flint activates just 2.5B of its 40B. The result: big-model quality at a fraction of the compute per token – though you still need enough memory to hold all the weights.

How to Choose the Right Open-Weight LLM

• Laptop or single GPU: small MoE models like Flint, or quantised versions of mid-sized models.
• Single server with multiple GPUs: mid-sized MoE models with ~40B active parameters.
• Cluster / enterprise: frontier-scale models like Mistral Large 4, served with vLLM, SGLang or a distributed stack like llm-d.
• Licence: “open-weight” isn’t always “open source” – check commercial-use terms before shipping.
• Benchmarks: test on your own tasks; leaderboard scores rarely match real workloads.

Running an Open-Weight Model Locally

With Ollama, trying a model takes two commands:

ollama pull <model-name>
ollama run <model-name> "Explain Kubernetes in one paragraph"

For production, use vLLM with an OpenAI-compatible endpoint:

pip install vllm
vllm serve <hf-org>/<model-name> --tensor-parallel-size 2

from openai import OpenAI
client = OpenAI(base_url="http://localhost:8000/v1", api_key="not-needed")
resp = client.chat.completions.create(
    model="<hf-org>/<model-name>",
    messages=[{"role": "user", "content": "Write a Python function to reverse a string"}],
)
print(resp.choices[0].message.content)

Security: Don’t Blindly Trust Open Models

Open weights give you control, but they also need scrutiny. Recent reports are a reminder:

• Cohere reported that some Qwen3.8 models occasionally send data to an Alibaba cloud bucket – so run models in a network-restricted environment and monitor egress.
• CrowdStrike and Booz Allen reported that Qwen3-Coder may write weaker code for certain prompts.
• The US DHS has banned DeepSeek across federal civilian agencies.

Best practice: run self-hosted models with no outbound internet access, use safetensors weights (never untrusted pickle files), and review AI-generated code like any other code.

Other AI Model News This Month

• Google opened its SynthID AI-content detector to the public.
• DeepSeek open-sourced DeepInfer (Apache 2.0), an inference engine for its own models.
• Claude Haiku 5.5 and Google Gemini 4 variants launched for API customers.

Conclusion

Open-weight LLMs are closing the gap with closed models, and MoE architectures are making them practical to run. Start small with a local model, benchmark on your own use case, lock down network access, and scale up with vLLM or Kubernetes when you’re ready.