mica-minicpm-v46-thinking

MiniCPM-V 4.6 Thinking — Mica Q4/Q8

Mica runtime artifacts for openbmb/MiniCPM-V-4.6-Thinking, pinned to revision 93d8f4b60ad5d1f763442cf4c19f2a71fa95af4a. The source and derivatives are Apache-2.0 licensed.

The runtime configuration, conversion policy, and validation evidence are maintained in the Mica Server repository. Validated artifacts are published together in the miguelamendez/mica-minicpm-v46-thinking repository, with manifests recording their source, size, and digest.

Model facts and provenance

The model combines a SigLIP2-400M vision encoder and a Qwen3.5-0.8B language model with mixed 4x/16x visual-token compression. The pinned text configuration declares 262,144 architectural positions. Upstream does not disclose a maximum training sequence, maximum input alone, or a separate hard output ceiling, so the architectural number must not be represented as a locally validated or training-safe context.

This is the long-chain-of-thought variant. Thinking is enabled by default and the pinned chat template permits it to be disabled. No low/medium/high effort levels or recommended reasoning-token budgets are disclosed.

For video, upstream defaults to at most 128 frames: approximately 1 FPS for short videos and uniform sampling for longer videos. Upstream examples request 512 new tokens for image work and 2,048 for video, but these are example request budgets rather than declared hard generation limits.

Mica artifacts and quantization policy

mlx/q4/    selective MLX affine Q4, group size 64
mlx/q8/    selective MLX affine Q8, group size 64
gguf/      Q4_K_M and Q8_0 plus full-precision projector
vllm/      not published; structural candidates failed the production gate

MLX quantizes eligible language projections while preserving the vision tower, vision mergers, recurrent linear-attention blocks, token embeddings, and output head. The effective checkpoints therefore average 13.274 bits per weight for Q4 and 14.222 for Q8; the directory names describe eligible affine layers, not blanket whole-model precision.

The portable Linux/NVIDIA comparison is AutoRound W4A16 group-128 versus GPTQ W4A16 group-128, followed by GPTQ W8A16 group-128. The vision tower, merger, recurrent linear-attention modules, token embeddings, and output head stay at source precision. A text-calibrated artifact is only an early candidate; it cannot be promoted until image/video calibration, real multimodal vLLM inference, and BF16 quality comparison pass. Blackwell-only formats are not in the current matrix.

Q4 and Q8 structural runs now calibrate from a real image and an ordered two-frame video. Both serialized checkpoints reload for batch-2 image and video generation. This is still not a quality artifact: the production calibration, BF16 comparison, and native Linux/NVIDIA vLLM tests remain mandatory.

No MTP drafter is included. Although the config declares mtp_num_hidden_layers, the published checkpoint contains no separable MTP or next-token tensors.

Current vLLM core recognizes this exact MiniCPM architecture. The earlier Apple vLLM-Metal 0.29 probe still fails during image-processor dispatch before loading weights; a newer core support entry does not establish Metal-plugin compatibility.

Mica validation boundary

MLX Q4 and Q8 passed real text, image, video, and OpenAI-compatible server inference. Q4 omitted a requested JSON suffix in one image test; Q8 matched the BF16 formatting and is preferred when structure fidelity matters. Recorded video peaks were 3.029 GB for Q4 and 3.176 GB for Q8. GGUF Q4_K_M and Q8_0 also passed ordered image/video inference.

The tests do not certify 262K context. Mica must count text, image slices, video frames, system/tool tokens, reasoning, and visible output within the selected profile's much smaller tested budget.

Validation records:

• docs/validation/minicpm-v46-thinking-mlx.md
• docs/validation/minicpm-v46-drafter-audit.md
• docs/validation/minicpm-vllm-quantization-macos.md
• docs/validation/gguf-runtime-macos-metal.md

Profile validation requirements

A MiniCPM profile must define text input, visual-token/media limits, maximum output, reasoning budget, KV precision, and concurrent sequences. It must warn above upstream-recommended or training context when those values become known, and fail above 262,144 total tokens or the active backend limit.