Step 3.7 Flash 196B-A11B

	Recommended use
Status	Supported on the qualified GGUF pipeline path — PP-2 serving, image input, structured output, and graphed draft chains; FP8 tuning is separate work
Best starting path	IQ4_XS trunk with the Q8_0 MTP head on two-card PP-2
Hardware	2× RTX PRO 6000 Blackwell
Use this when	The model does not fit one card and you want the qualified pipeline-parallel path

Use the Step cookbook. This is PP-2, not tensor parallelism. Topology and admission behavior are in serving.

Since the qualified pipeline landed, three more surfaces serve on the same endpoint:

• Image input (lane/step37-vision-20260830): the checkpoint's own perception_encoder tower runs in-engine behind MEMRA_STEP_VISION_DIR (default OFF; flags), with the vendor's exact tiling law. Per-token parity-gated against two independent references before any serving path; text requests stay byte-identical to the seam-off boot and keep MTP spec engaged. Receipts: research/step37-vision-20260830/.
• Structured output (lane/step37-postthink-grammar): response_format json_object/json_schema served post-think — the think phase runs unconstrained as trained, and the grammar clamps every token from the tokenizer's atomic think-close token (id 128799). Fail-closed named 400s if generation ends inside the reasoning channel. Receipts: research/step37-postthink-grammar-20260830/.
• Draft-graph serving (lane/step37-draft-graph-serving-20260830): the 3-head step-modulo draft chain is CUDA-graph captured on the serving shape, greedy and filtered-sampling arms byte-identical to eager. Receipts: research/step37-draft-graph-serving-20260830/.