AI Briefing
KO

Reka Edge 2603 support added

·2026.04.23 23:53

Key point

llama.cpp has added multimodal support for Reka Edge 2603 along with template fixes.

Details

Reka Edge 2603 multimodal support has been added to llama.cpp.

  • LLM: Llama-style, 32 layers, GQA 32/8 heads, RoPE, SiLU
  • Vision encoder: ConvNeXt V2-based 4-stage, depths [3, 3, 27, 3], GRN
  • Projector: 2-layer MLP(GELU), 2816 → 4096 → 4096, 64 visual tokens via adaptive 8×8 pooling per tile

The chat template includes a reasoning chunk, but Reka Edge 2603 itself does not support reasoning.

During the support work, the Jinja engine was also touched up. String repetition, ensure_ascii for tojson, and a no-op int() for integers were added to match Reka's template.

The author ran the VQA v2, RefCOCO A/B, and mobile-actions benchmarks along with llama-server tests, and stated that all changes are additive.

llama-cppreka
Read originalShare on Threads

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.

Related

HuggingFace BlogHugging Face Adds GGUF Model Support to TransformersRedditHugging Face Introduces Native GGUF Support in TransformersRedditDynamic Quantiser Released, Improving Accuracy Over Standard Quantization in VRAM-Constrained EnvironmentsNVIDIA BlogNVIDIA Unveils Local AI Acceleration and RTX Spark PCs at IFA 2026HuggingFace BlogMeta Releases Muse GlimmerHacker NewsShow HN: Reame – a CPU inference server that gets faster the more it runs