Repository navigation
Commit e7d2a4d
fix: replace retired Qwen2.5 model in tutorial 30
Tutorial 30 calls `Qwen/Qwen2.5-7B-Instruct` through `provider="together"`,
which routes to Together's `Qwen2.5-7B-Instruct-Turbo`. Together retired
that model from serverless, so the HF provider mapping now reports
`status=error` and the model is gone from the router's model list. The
nightly log already shows the health-check warning for it.
Swap in `Qwen/Qwen3.5-9B` and disable thinking. Qwen3.5 is a hybrid
thinking model, and left on it spends ~8x the completion tokens (1292 vs
165 on this tutorial's prompt) and prepends a 4 KB reasoning trace, which
would make the tutorial's sample response unrepresentative.
Note this does not by itself make the nightly green: that run failed with
`402 Payment Required` — "You have depleted your monthly included
credits" — which no code change can fix.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>1 parent a7cae0f commit e7d2a4d
1 file changed
Lines changed: 4 additions & 3 deletions
| Original file line number | Diff line number | Diff line change | |
|---|---|---|---|
| |||
312 | 312 | | |
313 | 313 | | |
314 | 314 | | |
315 | | - | |
| 315 | + | |
316 | 316 | | |
317 | 317 | | |
318 | 318 | | |
| |||
409 | 409 | | |
410 | 410 | | |
411 | 411 | | |
412 | | - | |
| 412 | + | |
| 413 | + | |
413 | 414 | | |
414 | 415 | | |
415 | 416 | | |
| |||
448 | 449 | | |
449 | 450 | | |
450 | 451 | | |
451 | | - | |
| 452 | + | |
452 | 453 | | |
453 | 454 | | |
454 | 455 | | |
| |||
0 commit comments