Open-weight model

Open-weight models are large language models (or other machine learning models) whose trained weights are published for anyone to download and run. This is distinct from open-source, which would also require the training data and code used to produce the model to be public — most open-weight models do not meet that bar, since labs frequently withhold the data and training recipes even when releasing the weights themselves.

Licensing terms vary between open-weight models. Some permit unrestricted commercial use (eg. Apache 2.0, MIT). Others carry custom licenses that restrict commercial use, redistribution, or use above a certain scale (eg. Meta’s Llama community license).

Hugging Face is the primary hub for discovering and downloading open-weight models, hosting over two million models contributed by individuals, research labs, and major organizations. The other major resource in this space is Ollama’s model database, which is a more curated library of models that are proven to be stable with the Ollama model manager (see model access layer).

There are pros and cons to running open-weight models on your own hardware:

  • Free (just the cost of the hardware and electricity).
  • Highly customizable.
  • Secure and private.
  • Zero network latency.
  • Work offline (no internet connection required).
  • High availability. Can be run 24/7/365.
  • Six to 12 months behind the frontier models in capability.
  • Require a lot of RAM and GPU VRAM to run large models.

Hardware requirements

It is useful to think of open-weight models as falling into a spectrum of size tiers, with each tier having different minimal hardware requirements:

  • Ultra-light (1–3B): Run on almost any machine with 8 GB+ RAM. Capable of basic chat and simple tasks.
  • Small (3–8B): A good balance of quality and speed. Need ~8–16 GB RAM. The recommended starting point for most local setups.
  • Medium (8–13B): Well-rounded general-purpose models. Run well on 16–24 GB RAM with a mid-range GPU.
  • Large (30–70B): High capability. Require at least 32 GB RAM and a dedicated GPU with 24+ GB VRAM.
  • Very large (70B+): Near-frontier capability. Require high-end workstations with multi-GPU setups and at least 64 GB each of RAM and VRAM.

To estimate whether your own hardware can run a given model — accounting for quantization level (Q4–F16), context length, and expected token throughput (TPS) — try the Local AI VRAM Calculator & GPU Planner. It maps a hardware setup and use case to model recommendations, with the caveat that the figures are planning estimates, not performance guarantees.

Custom hardware like the Nvidia DGX Spark give you the ability to run and train large language models on your own hardware.

Open-weight model families

As of 2026, some notable open-weight model families include (several use a mixture-of-experts (MoE) architecture):

  • Llama models from Meta. The Llama family did much to kick-start the open-weight era, and its widespread adoption has made the underlying architecture a de facto reference point that many other open models build on. The Scout variant is known for its massive 10-million token context window. Released under Meta’s community license, which permits commercial use but imposes some restrictions (including for very large deployments).
  • Gemma models from Google. These are Google’s open-weight counterparts to its proprietary, frontier Gemini models. Gemma models are optimized to run on consumer hardware.
  • Qwen models from Alibaba Cloud are available in a wide range of sizes, so there’ll always be one variant that you can run on your hardware.
  • Mistral models from Mistral AI, a French company that produces both open-weight models and proprietary models accessible via web service APIs. Their Mixtral models helped popularize the mixture-of-experts (MoE) architecture in open models.
  • Phi models from Microsoft. A family of small language models (SLMs) that punch well above their size on reasoning and maths, achieved largely by training on high-quality, "textbook-quality" filtered and synthetic data rather than sheer scale. Optimized for on-device and edge use, and released under the permissive MIT license. Recent releases include Phi-4 (14B) and the compact Phi-4-mini, with multimodal variants.
  • gpt-oss models from OpenAI. gpt-oss-120b and gpt-oss-20b, released under the permissive Apache 2.0 license in August 2025, were OpenAI’s first open-weight models since GPT-2.
  • Kimi models from Moonshot AI. Kimi K2.5 (released January 2026) is an open-weight mixture-of-experts model with 1T total parameters (32B activated per token) and a 256K-token context window. It matches frontier models on several benchmarks and is multimodal across text, code, image, and video (understanding only, no generation). Kimi 2.5 leads in mathematical reasoning and is optimized for "agent swarming".
  • GLM models from Z.ai (formerly Zhipu AI). The GLM-4.5 and GLM-4.6 releases are strong open-weight MoE models, well-regarded for their "interleaved thinking" and demonstrated capabilities in both agentic and vibe coding workflows.
  • DeepSeek models from DeepSeek AI. This AI lab made waves in early 2025 when DeepSeek-R1 demonstrated near-frontier reasoning performance at a fraction of the training compute cost that had previously been assumed was necessary to develop frontier models. Its "engram" memory system has been proven to maintain high coding performance across 1M+ token contexts.
  • MiniMax models from MiniMax AI, which also develops closed-weight models. MiniMax M2.5 is well-regarded, posting near-frontier agentic and coding scores (~80% on SWE-Bench Verified) at an exceptionally low cost.

Nvidia continues to expand its open-weight offerings, most recently with the Nemotron 3 family (Nano, Super, and Ultra variants).

Open-weight coding models

The following open-weight models are highly regarded for programming tasks. Small models (≤ 8B) run comfortably on a laptop or CPU-only setup. Larger models require a GPU with ≥ 16–24 GB VRAM.

  • Qwen2.5-Coder 32B Instruct: Alibaba’s dedicated coding model that consistently tops open-weight coding benchmarks. Strong across Python, JavaScript, Go, and SQL. Supports a 32K context window. Apache 2.0 license.
  • DeepSeek-Coder-V2 Instruct: A Mixture-of-Experts model (236B total, 21B active) with very strong multi-language coding performance and a 128K context window. One of the most capable open-weight options for complex, multi-file tasks. MIT license.
  • Codestral 22B: Mistral AI’s dedicated code model, optimized for low-latency code completion and fill-in-the-middle (FIM) tasks. Well suited to IDE integration. Supports 32K context. Mistral AI Non-Production License (commercial use requires a separate agreement).
  • StarCoder2 15B: The successor to StarCoder, trained on The Stack v2. Strong on Python, JavaScript, and C/C++, and well suited to code search and autocompletion tasks. Apache 2.0 license.
  • Code Llama 70B Instruct: Meta’s instruction-tuned coding variant of Llama 2, with a 16K context window and strong performance on multi-language tasks and large-scale refactoring. Meta’s non-commercial license restricts use to research.

See also