Providers / meta-llama/llama-3.3-70b-instruct
meta-llama/llama-3.3-70b-instruct as each provider declares it
5 of the 11 endpoints declare at least one material setting that differs from the official reference. Declared values, not measured behaviour.
Reference. Context 131,072 tokens, from unsloth/Llama-3.3-70B-Instruct config.json at 99cd0d2c82 (max_position_embeddings). Quantization: no publicly verifiable reference.
Meta's repository (meta-llama/Llama-3.3-70B-Instruct) is gated and cannot be read anonymously, so the context length is read from the unsloth mirror of the same weights.
Meta publishes no reference quantization that can be read anonymously (its FP8 repository answers 401 to an anonymous request, which does not say whether it exists), so the quantization each provider declares is shown and not compared.
Provider rows: OpenRouter providers page, the endpoint JSON. Read 2026-10-19.
Endpoints
| Provider | Quantization | Context | Max output | Tools / reasoning | Differs from the reference |
|---|---|---|---|---|---|
| AkashML akashml/fp8 | fp8 (not compared) | 131,072 | 128,000 | yes / n/a | none declared |
| Cloudflare cloudflare/fp8 | fp8 (not compared) | 24,000 | 21,600 | no / n/a | declares a 24,000-token context; the reference config says 131,072 (18%)declares no tools parameter; the official model supports tool calling |
| CoreWeave coreweave/fp16 | fp16 (not compared) | 128,000 | 115,200 | yes / n/a | declares a 128,000-token context; the reference config says 131,072 (98%) |
| DeepInfra deepinfra/turbo | fp8 (not compared) | 131,072 | 16,384 | yes / n/a | none declared |
| Google google-vertex | none (not compared) | 128,000 | 115,200 | no / n/a | declares a 128,000-token context; the reference config says 131,072 (98%)declares no tools parameter; the official model supports tool calling |
| Google google-vertex/us-central1 | none (not compared) | 128,000 | 8,192 | yes / n/a | declares a 128,000-token context; the reference config says 131,072 (98%) |
| Groq groq | none (not compared) | 131,072 | 32,768 | yes / n/a | none declared |
| Novita novita/bf16 | bf16 (not compared) | 12,288 | 11,059 | yes / n/a | declares a 12,288-token context; the reference config says 131,072 (9%) |
| Parasail parasail/fp8 | fp8 (not compared) | 131,072 | 16,384 | no / n/a | declares no tools parameter; the official model supports tool calling |
| SambaNova sambanova-turbo | none (not compared) | 131,072 | 3,072 | no / n/a | declares no tools parameter; the official model supports tool callingdeclares at most 3,072 output tokens |
| Together together | none (not compared) | 131,072 | 2,048 | yes / n/a | declares at most 2,048 output tokens |
Changes since tracking began
None yet. A row is recorded only when a declared value or the reference moves.
How this is read
Once a day the tracker reads OpenRouter's public endpoint list for each model (one row per provider deployment: declared context length, declared maximum output, declared quantization, and the request parameters it accepts) and the official repository's config.json at its head commit. Both reads are anonymous. A new row is recorded only when either moved, so the history below is a list of real changes.
Every value in the tables is what a provider declares to OpenRouter. No request is sent to any model, so nothing here says how an endpoint behaves: one that declares a 131,072-token context may still cut off earlier, and one that declares the official quantization may still serve something else.
A quantization the provider does not declare cannot be compared, and is counted as undisclosed rather than as matching. A quantization of higher precision than the official weights is shown and not counted as a difference.
Material means one of: a declared context under 95% of the reference config, a declared quantization other than the official release, no tools parameter on a model that supports tool calling, no reasoning control on a model that exposes one, or a reasoning model capped under 8,192 output tokens. A context between 95% and 100% of the reference is listed and not counted as material.
Sampling defaults, chat templates and stop tokens are not in the endpoint metadata and are not compared here. Moonshot's K2 Vendor Verifier (github.com/MoonshotAI/K2-Vendor-Verifier) measures tool-call behaviour across providers for Kimi K2 by sending requests; this page does not.
OpenRouter lists only the providers that serve through it. Nothing here ranks providers, and a provider with no listed difference has not been shown to match the reference, only to declare nothing that differs.