internvl2.5-2b-crispembed-GGUF

InternVL2.5-2B — CrispEmbed GGUF

GGUF conversions of OpenGVLab/InternVL2_5-2B for use with CrispEmbed.

Model Details

Available Quantizations

Note: Vision encoder weights are kept at Q8_0 minimum even in Q4_K to preserve OCR accuracy. The Q4_K savings come primarily from the LLM decoder.

Usage with CrispEmbed

#include "crispembed.h"

// Auto-detects InternVL2 architecture from GGUF metadata
void *ctx = crispembed_math_ocr_init("internvl2.5-2b-q4_k.gguf", 4);

int len;
const char *text = crispembed_math_ocr_recognize(ctx, pixels, w, h, channels, &len);
printf("%s\n", text);

crispembed_math_ocr_free(ctx);

from crispembed import CrispMathOcr

ocr = CrispMathOcr("internvl2.5-2b-q4_k.gguf")
text = ocr.recognize("document.png")

Parity Verification

All components verified against the Python reference implementation:

Conversion

Converted using models/convert-internvl2-to-gguf.py from CrispEmbed:

python models/convert-internvl2-to-gguf.py \
    --model OpenGVLab/InternVL2_5-2B \
    --output internvl2.5-2b-f16.gguf --dtype f16

# Then quantize with the C++ quantizer:
./crispembed-quantize internvl2.5-2b-f16.gguf internvl2.5-2b-q8_0.gguf q8_0
./crispembed-quantize internvl2.5-2b-f16.gguf internvl2.5-2b-q4_k.gguf q4_k

Architecture

Image (448x448 per tile, 1-12 tiles)
  → Conv2D patch embed (14x14, stride 14) → 1024 patches
  → Prepend CLS + position embedding
  → 24x InternViT blocks (LayerNorm → MHSA → LayerScale → residual
                           LayerNorm → GELU MLP → LayerScale → residual)
  → Remove CLS → pixel unshuffle (4:1, 1024→256 tokens, dim 1024→4096)
  → LayerNorm → Linear(4096→2048) → GELU → Linear(2048→2048)
  → Splice into text token sequence
  → 24x InternLM2.5 blocks (RMSNorm → GQA(16/8) + RoPE → residual
                              RMSNorm → SwiGLU FFN → residual)
  → RMSNorm → LM head → logits → greedy decode

Credits

• Original model: OpenGVLab/InternVL2_5-2B (MIT License)
• GGUF conversion: CrispEmbed

Provenance and EU AI Act Art. 53 note

• Upstream model: OpenGVLab/InternVL2_5-2B — published by OpenGVLab.
• Upstream licence: mit. This repository redistributes under the same terms; it grants no rights the upstream licence does not.
• What was done here: format conversion and/or quantisation only (GGUF). No training, no fine-tuning, no merging, no distillation, no change to architecture, vocabulary or capability. Only the numeric representation of the upstream weights differs.
• Training data: documented — where it is documented at all — by the upstream provider; see the upstream model card. No training data was used, added or selected by this repository.
• Provider status: under Regulation (EU) 2024/1689 the upstream authors remain the provider of this model. Converting the serialisation format does not make this repository the provider of a new general-purpose AI model, and no such claim is made. Questions about training content, copyright policy or model capability belong upstream.