What happened
The Qwen3.5-4B-based v0.4 weights first appeared on October 29. A November 2 revision removed the vision encoder without changing language weights or answers, according to the repository. The November 4 documentation adds evaluation details and an original benchmark chart; this is not a newly trained November 4 base model.
The Apache-2.0 release answers typed questions in one forward pass per question instead of generating prose. Its current text-only weights occupy 8.41 GB. Correct operation requires the bundled loader’s attention masks; ordinary causal serving changes its behavior. English is the only evaluated language.
Macrocosmos reports a 78.6 JevBench composite, attributed to Benchmark Heaven’s November 4 measurement. The card says its newer question draw has not yet been equated with the public board, so direct ranking claims need caution. It also discloses checkpoint selection on a public test split and text overlap in another evaluation.
Why it matters
Structured decision models could reduce unnecessary text generation in routing and classification. Their confidence scores still need testing on the intended workload. Here, serving compatibility and test-set comparability are as important as the headline score; neither low latency nor a benchmark position guarantees reliable business decisions.
What to watch
Watch for the equated public leaderboard entry, independent application tests and support for common serving engines.
Sources & verification
Source links are not a count of independent confirmations.
01Report a correction
Help us get it right
Include the disputed statement and, if possible, a supporting source URL. Please do not send personal or sensitive information.
