DeepSeek V4's biggest problem was never the model.
It was the harness. A GitHub dev built JSpace, a wrapper that changes ZERO weights. It just manages how the model works. Same model. Same brain. Scores jumped across every benchmark. → DeepSWE: 62.7 to 72.0 → NL2Repo: 61.5
What are you doing this weekend?
Wrong. If you're into LLMs, you're installing ollama then trying out Intersignal Braid v1.5.2: let your AIs *share state* over the LAN or WiFi. Download for Mac, Windows, and Linux:
DeepSeek-v4-Flash-Vision is better than Opus 4.8
From my limited testing I will put ox alpha somewhere between qwen3.8 27b and deepseek v4 pro.
If it is really a flash model, that’s great!
You have reached the end of the archive
All of deepseek