Changes / Feeds / Qwen/Qwen3-8B-FP8
Qwen/Qwen3-8B-FP8
2 changes to a tracked key since 2025-04-28, newest first. Head as last read: 220b46e3b2180893580a4454f21f22d3ebb187d3.
A quantization of Qwen/Qwen3-8B: 252 layers re-encoded; of the 147 tensors left unquantized, 147 byte-identical to the base. Compared tensor by tensor from sha256 fingerprints: 220b46e3b218 against Qwen/Qwen3-8B at b968826d9c46. A hash is equal or not; it says nothing about what an edit does.
| Date | Advisory | What changed, and what it does |
|---|---|---|
| 2025-05-19 | x_NXMC-2025-0122 | the chat template text changed but rendered byte-identically on the three fixed conversations (system+user, multi-turn, tool call), all of which end on a user turn; a conversation ending on an assistant turn (training data, prefill) was not rendered (2025-05-19).02d96047e692 -> a8f74dfface5 Open in app |
| 2025-04-28 | x_NXMC-2025-0121 breaking | EOS stop list 151645 to [151645,151643]; generation now also stops on 151643, and still stops on every token it stopped on before; generation_config.json BOS declared as 151643 (was unset); generation defaults changed (do_sample unset to true, temperature unset to 0.6, top_k unset to 20, top_p unset to 0.95); a call that does not pass these explicitly now generates differently (2025-04-28).18a569c78b18 -> add3e8dee80b Open in app |
Pin it, and let a bot move the pin
from_pretrained("Qwen/Qwen3-8B-FP8", revision="220b46e3b2180893580a4454f21f22d3ebb187d3")
With the Renovate preset, a pull request moves this pin when the head moves, and links back here.
[](https://www.neurarch.com/changes-feeds/models/Qwen/Qwen3-8B-FP8.html)