I'm testing the 128k Claw-Scout-32B model via their API. The docs claim a 128k context window. My system prompt explicitly sets `max_tokens` to 1000 and I use their `context_window` parameter in the request.
It accepts a 90k token prompt without error. The response is coherent but clearly only references material from the first ~4k tokens. The rest is ignored. No warning, no truncation error, just silent failure.
My config is straightforward:
```
model: "claw-scout-32b"
max_tokens: 1000
context_window: 128000
temperature: 0.2
```
Has anyone else validated the actual working context? I suspect the advertised limit is for a different, more expensive inference mode. Their SLA is useless for this—it only covers uptime, not functional accuracy. What's the real limit before silent degradation?
read the fine print
Yeah, we've seen similar behavior with their earlier models. The `context_window` parameter often feels like a suggestion to their load balancer, not a hard limit enforced during inference.
I'd try two things:
1. Split your prompt. Send a system prompt first that asks the model to confirm receipt of a unique marker placed at the very end of your main prompt. If it can't see the marker, you've got your answer.
2. Check the token count they bill you for. Their API sometimes returns a different `usage.prompt_tokens` than what you'd expect from your own tokenizer. If it's way lower, they're truncating upstream.
Silent truncation is the worst. Makes you question every summarization or RAG task. Their support just links to the docs, but maybe we can crowd-source a real limit.
You've basically described the benchmark for half the "long context" models out there right now. The silent degradation is the feature, not the bug. They'll happily charge you for the 90k token input, then let the model operate on a tiny effective window.
You should check the `usage.prompt_tokens` in the response. I'd bet a coffee it's nowhere near 90k. They're likely chunking it or applying a sliding window at the inference layer and just not telling you. The advertised limit is usually for the *input*, not the model's functional attention across that input.
cg