Open-Weight Models
Choose open-weight models without guessing the license, hardware, or serving path.
Pick the model you can actually run, adapt, and defend. Open-weight models make trained weights available, but the license still decides what you can run, adapt, or redistribute. Compare model families by the parts teams actually ship: license terms, weight access, inference stack, data boundary, eval plan, and operating cost. When the shortlist is no longer vague, K3 Nova is the practical next stop for trying the model workflow instead of just reading about it.
Updated Aug 4, 2026
Source-backed notes
Independent planner
Direct answer
What are open-weight models?
An open-weight model is an AI model whose trained parameters are available to download or inspect under stated terms. That is useful because you can run, adapt, evaluate, or host the model outside the original vendor's API path.
The label is not a blanket permission slip. Training data, training code, benchmark harnesses, commercial service rights, attribution rules, and acceptable-use policies may still be closed or restricted.
| Term | What it usually means | What to verify |
|---|---|---|
| Open-weight | Trained parameters are available to download, inspect, run, or adapt under stated terms. | Read the model card, license, acceptable-use policy, tokenizer files, and redistribution terms. |
| Open source AI | More of the system is available, such as code, data information, training recipes, and reproducibility materials. | Do not assume weights alone satisfy open-source obligations for your legal or compliance review. |
| API-only | You can call a hosted model, but cannot download the trained weights or operate the model yourself. | Review data retention, region, logging, rate limits, and vendor continuity instead of hardware fit. |
Why this exists
Open weights shift the decision from access to stewardship.
Once weights are available, the real work moves to license review, runtime fit, prompt format, quantization, safety testing, observability, and support boundaries. The homepage is built as a planner because teams need a decision surface, not another generic definition page.
Current shortlist
Six open-weight paths worth checking before you pick a runner.
Readers usually need more than a name list. Use this table to decide what deserves a real test, then move the candidate into K3 Nova when the license and runtime questions are clear enough to evaluate on product behavior.
| Family | First fit question | License or runtime checkpoint | Natural next test |
|---|---|---|---|
| gpt-oss120b / 20b | Do you need a permissive reasoning model that can run under your own infra controls? | Confirm memory target, tokenizer, safety policy, and support boundary before production. | Use K3 Nova to test reasoning depth, tool calls, and structured output on your own task traces. |
| Qwen3MoE and dense sizes | Is multilingual, coding, or agent throughput more important than a single benchmark score? | Check exact release, context behavior, and serving framework support for the chosen size. | Use K3 Nova to compare prompt formats, locale behavior, and code-edit quality side by side. |
| LlamaLlama 4 Scout / Maverick | Do you need the Llama ecosystem enough to accept gated access and custom license review? | Record license acceptance, attribution, redistribution, and acceptable-use obligations. | Use K3 Nova once legal review says the product flow can safely include a Llama-family candidate. |
| MistralMistral 3 family | Do you need a permissive model family with enterprise deployment options? | Verify the exact model card because Mistral license posture varies by model. | Use K3 Nova to test whether the smaller, faster variant handles the workflow before scaling up. |
| Kimi K32.8T-class MoE | Are frontier-style agent and long-context gains worth custom commercial license review? | Read the Kimi K3 License before benchmarking because commercial service scale can change the approval path. | Use K3 Nova after license review to see whether Kimi K3 improves the real agent loop, not just a benchmark row. |
| DeepSeek V4Pro / Flash | Do you need million-token context or low-cost routing more than vendor ecosystem fit? | Validate encoding, context budget, provider terms, and safety posture for the exact mode. | Use K3 Nova to test long-context traces, latency, and fallback routing before committing infrastructure. |
Current family snapshot
Model names are not enough. Compare release terms and run paths.
| Family | License posture | Run path | Best fit |
|---|---|---|---|
| gpt-oss120b / 20b | Apache 2.0 plus usage policy | Self-managed or partner-hosted; not OpenAI API or ChatGPT | Reasoning, tool use, structured outputs, private infrastructure experiments |
| Qwen3MoE and dense sizes | Apache 2.0 for Qwen3 open-weight releases | SGLang, vLLM, Ollama, LM Studio, MLX, llama.cpp, KTransformers | Multilingual work, coding, agent workflows, long-context experiments |
| LlamaLlama 4 Scout / Maverick | Llama community license | Meta utilities, transformers, multi-GPU inference for large variants | Broad ecosystem support with license acceptance and attribution review |
| MistralMistral 3 family | Apache 2.0 for many open models; model cards remain authoritative | vLLM, TensorRT-LLM, SGLang, managed clouds, local edge variants | Enterprise-friendly permissive releases, multilingual and multimodal paths |
| Kimi K32.8T-class MoE | Kimi K3 License with commercial-service and attribution clauses | vLLM, SGLang, TokenSpeed, plus hosted Kimi API access | Agentic coding, long-context knowledge work, and native multimodal experiments |
| DeepSeek V4Pro / Flash | MIT for the repository and model weights | Local deployment recipes, DeepSeek API, and provider-hosted paths | Long-context reasoning, coding, and cost-sensitive high-throughput workloads |
Common questions
Use the page in the order your decision will happen.
A good open-weight decision moves from vocabulary to evidence, then to a small test. These are the four jobs this homepage is meant to answer without making you decode a vendor announcement first.
I need the definition.
Start with weights: can a practitioner obtain the actual model parameters, and what rights come with them?
I need a model list.
Compare families by license, run path, and fit. The same name can point to different sizes, formats, or terms.
I need license confidence.
Separate Apache-style grants, custom community licenses, gated downloads, naming rules, and commercial thresholds.
I need to test one.
Once the shortlist has evidence, K3 Nova is the next practical place to test behavior in a real product workflow.
Start with the boring checks
The winning model is the one your team can operate.
Use this page to keep the license, runtime, and data-boundary questions honest. When those questions are clear enough to test, move the work into K3 Nova and see how the model behaves in a real product flow.