PR #15683: server: preserve thinking in /api/generate and populate parameter_size in /api/tags for safetensors

• Repository: ollama/ollama
• Author: serenposh
• State: open | merged: False
• Link: https://github.com/ollama/ollama/pull/15683

Description (problem / solution / changelog)

Summary

Fixes two independent bugs surfaced on gemma4:26b-mxfp8 / gemma4:26b-nvfp4.

1. /api/generate silently drops thinking for models that think by default (#15681)

GenerateHandler initialized the builtin parser before the capability-gated default for req.Think was applied. Parsers that gate thinking output on the value passed to Init — notably Gemma4Parser, which has an explicit // When thinking is disabled, silently discard channel content branch — therefore saw thinkValue == nil and dropped the reasoning, even though the model was emitting it (visible via a large eval_count but short response).

Moving the capability check + default above the parser Init call so the parser sees the resolved req.Think value. This matches ChatHandler, which already does the two steps in the correct order — and is why /api/chat / /v1/chat/completions return reasoning correctly on the same model.

Callers that explicitly set think: false are unaffected — the default only kicks in when req.Think == nil.

2. /api/tags returns empty parameter_size for safetensors models (#15679)

ListHandler populated Details purely from the manifest's ConfigV2, whose ModelType / FileType are not written for safetensors models during create. /api/show already works around this by reading the safetensors headers via xserver.GetSafetensorsLLMInfo / GetSafetensorsDtype; mirror the same enrichment in ListHandler so the two endpoints stay consistent.

Note: the separate observation in #15679 that the reported count (e.g. 8.7B) is the active-parameter count for MoE variants rather than a "26B-A4B"-style total is a deeper metadata question — out of scope for this PR. This change at minimum stops /api/tags from returning an empty string and makes it match the existing /api/show value.

Verified locally

• go vet ./server/ — clean
• go build ./server/ — clean
• go test ./server/ — all pass (2.5s)
• go test ./model/parsers/ — all pass

Needs manual verification by reviewer

Neither author has a machine with gemma4:26b-mxfp8 / a safetensors model available, so these runtime checks weren't performed:

• curl /api/generate -d '{"model":"gemma4:26b-mxfp8","prompt":"Moin","stream":false}' → response now includes populated thinking field.
• Same request with "think": false → thinking empty (no regression for explicit opt-out).
• curl /api/generate -d '{"model":"llama3.2","prompt":"hi"}' (non-thinking model) → unchanged behaviour.
• curl /api/tags on a machine with a safetensors model → details.parameter_size is populated (matches /api/show).
• curl /api/tags on a machine with GGUF-only models → unchanged (manifest config still used).

Changed files

• server/routes.go (modified, +40/-16)