Cohere Releases North Micro: 2.4B Open-Weight Vision-Language Model
pmttyji · reddit · 2026-08-13
CohereLabs has released North-Micro-Vision-Instruct, a 2.4B-parameter open-weight vision-language model under the Apache 2.0 license.
Key Details:
- Architecture: Features a 2B language model backbone (128K context) and a 400M vision encoder (custom-trained from SigLIP 2 SO400M).
- Capabilities: Supports native-resolution image processing, VQA, captioning, grounding, OCR, chart/document understanding, multilingual inputs, and multi-image support.
- Limitations: Multimodal context validated only up to 8K tokens; no tool calling/agentic workflows; system prompts not recommended. Positioned as a compact foundation for customization rather than a general-purpose assistant.
Related event: Cohere Open-Sources 2.4B Vision-Language Model North Micro(4 posts)→
More from Models
- DeepSeek V4 Pro Reported to Prematurely Halt in Agentic Coding — karminski3 · 2026-08-13
- xAI Launches Grok 4.6: Major Performance Boost at Same Price, Now in Hermes Agent — NousResearch · 2026-08-13
- Visualizing Benchmarks: Qwen 3.8-Max Outperforms Opus 4.8 Across Multiple Metrics — deliprao · 2026-08-13
- Grok 4.6 tested on bug bench: outperforms predecessor, becomes new default — PawelHuryn · 2026-08-13
- Grok 4.6 tops AI models under $5 workload budget, showing high intelligence per dollar — XFreeze · 2026-08-13
- OpenAI and Anthropic Models Dominate in Long-Running Autonomous Workflows — scaling01 · 2026-08-13