Replaces the v0 baseline, which was the UNMODIFIED Qwen3.5-4B and stays fetchable at the pre-v1 tag.
invented-identifier 0.378 -> 0.023, panel-reading 0.595 -> 0.967, ScreenSpot 0.718 -> 0.907, vision_probe 0.727 -> 0.732 (no regression), VizWiz false-refusal 0.220 -> 0.227 (flat).
Read the caveats in the card: much of the ScreenSpot gain is learning the normalised 0-1000 coordinate convention rather than better sight, and v1 does not read photographs better than v0 -- it merely stops over-refusing.
v0 has a vision baseline now: vision_probe 0.727 +/- 0.030
The card and recipe both said no vision benchmark existed. One does - a curated
slice mined from MMStar, RealWorldQA and V*Bench, selected model-independently
and scored in its own lane. v0 measured 0.727 +/- 0.030 over 220 items with a
clean contamination canary, on the untrained base.
The private camera task is referenced without its score: it runs only against
local services and those frames stay on our hardware.
Correct the card and recipe: v0 weights ARE published here
README's Versions section and recipe.json still said the weights were not
mirrored, contradicting the Weights section and publication.json after
commit 8f16249 uploaded them. Both shard SHA-256 digests match
Qwen/Qwen3.5-4B @ 851bf6e8, so the mirror claim is verified, not asserted.
The weights are mirrored here now, so the repository is self-contained rather
than a card pointing somewhere else. Adds a working load snippet and a serve
line carrying the two settings this model needs and will not announce: the
qwen3_xml tool parser, and thinking disabled.
Records the weights approval alongside the original scaffolding one instead of
overwriting it. The gate on TRAINED checkpoints stays closed.
v0 weights: Qwen3.5-4B BF16 at pinned revision 851bf6e8
v0 is the starting line: the pinned base at BF16, measured on our own hardware
so every later version has an exact thing to beat. Records throughput, the
vision check, the tool-call parse and both canaries.
The throughput is the notable result - at BF16 this 4B runs at about the same
single-stream rate as the 27B NVFP4 on the same box, which is a memory-bandwidth
finding and the case for quantizing the deployment lane.
The chat-template gate was right but did not finish the argument: more than one
base passed it. Among those, Qwen3.5-4B shares Qwen3_5ForConditionalGeneration
with the resident local 27B vision model, so it needs no second serving path,
and it has 4 KV heads against 8 for half the KV cache. The vision tower
configuration is identical, so no sight is traded.
Records the two carried differences: the non-JSON call serialization (which the
control run must report rather than hide) and thinking-on-by-default with its
required empty-content canary.
Scaffolding: base pinned, identity scheme defined, no run executed
