The 1M context claim, examined
The model card and the API catalogue disagree about how much context you can actually buy. Both figures are below, because both are true in different senses.
Inkling 2 Small's context window is 1M tokens according to the official model card. Through Tinker, however, the model is served at 64K and 256K tokens. The 1M figure describes what the architecture supports; the smaller figures are what you can currently send in a request.
Advertised against served
| Figure | Tokens | Source | Status |
|---|---|---|---|
| Advertised maximum | 1,000,000 | Model card | Model capability |
| Tinker variant | 65,536 | Tinker catalogue | Available now |
| Tinker variant | 262,144 | Tinker catalogue | Available now |
| Training, phase 2 | 524,288 | Announcement | Not a limit |
| Training, phase 3 | 1,048,576 | Announcement | Not a limit |
| Artificial Analysis reading | 1,000,000 | Artificial Analysis | Independent |
Why the numbers differ
A context window is an architectural capability: how many tokens the attention mechanism can address before positional generalisation breaks down. A served limit is a deployment decision: how much KV-cache memory the vendor is willing to allocate per request at a given price.
Thinking Machines trained at 512K and 1M context in their later phases and evaluated with a 1M-token maximum trajectory limit, so the long-context ability is real. But the Tinker catalogue exposes 64K and 256K endpoints. Reaching 1M context means either a config change on the vendor side or self-hosting — and the latter needs the weights, which are not public yet.
Both served variants are priced identically for inference, so the 256K variant is strictly better value if your workload fits within it. Only fine-tuning differs: $3 per million tokens at 256K against $1.5 at 64K.
What long context costs in memory
KV cache grows linearly with context length. The model card does not publish head counts, so an exact figure is not derivable; the hardware page states its assumption openly and lets you scale it. Roughly: at a 32K working context the cache is a few gigabytes, and a 256K context multiplies that by eight.
Practical guidance
- Most RAG and agent workloads fit comfortably in 64K
- Long-document work should use the 256K variant
- Do not plan around 1M until it appears in the catalogue or the weights ship