Long-context model tiers
Explore open-weight models in published 1M, 256K, and 128K context tiers, with sources and caveats about memory, configuration, and useful context.
- What is included
- Selected models with a context tier recorded in our sourced directory. Profiles marked Under review are excluded. Optional context extensions are not promoted into a higher tier.
- How this list is ordered
- Largest published context tier first, then model name within a tier. K/M labels are approximate groups; they do not assert equal exact token limits or equal long-context performance.
Publisher sources are linked for every model. No independent performance scores have been measured. On smaller screens, scroll the table horizontally.
| Model | Published context | Deployment note |
|---|---|---|
| 1M | Separate prompt processing and generation tests for this architecture.Compare model | |
| 1M | The large checkpoint requires its own distributed-serving evaluation.Compare model | |
| 1M | A large multimodal checkpoint: test memory at your intended input length and concurrency.Compare model | |
| 1M | Validate the sparse-attention backend as well as the context setting.Compare model | |
| 1M | This profile covers NVFP4 weights; verify the supported runtime and hardware recipe.Compare model | |
| 256K | Test text, image, and audio workloads independently.Compare model | |
| 256K | Use representative repository context and tool traces, not only synthetic text.Compare model | |
| 256K native | Listed at its native 256K tier; the optional extension is a separate configuration.Compare model | |
| 128K | The context ceiling is not a promise of on-device memory capacity.Compare model |
Capacity is not retrieval quality
A supported token limit describes an input envelope. It does not establish that a model reliably uses every part of a document, retrieves the right evidence, or responds within your latency target. A useful evaluation asks answerable questions at different document positions and scores whether the response cites the relevant information.
Compare the same amount of application work
Tokenizers can produce different lengths for the same document. Keep the source documents and requested answers consistent, while recording each model's actual token counts. Measure short, typical, and long inputs separately. Do not infer long-context throughput from a short-prompt run.
Check the exact limit upstream
This table groups publisher labels into approximate context tiers. Native and extended limits, modality processing, and deployment settings may differ. Open the linked model card before setting a server limit, and allow for generated output within the supported sequence budget.
Before you choose hardware
KV cache and context length: why a model runs out of VRAM explains the next planning step. For the distinction between publisher specifications, calculations, and measurements, read the BenchGrid methodology.