Sabitlenmiş Tweet

🚨 Hermes 4 — the massive new model from @NousResearch
Three sizes (14B, 70B, 405B). Open-weight. And on paper, it already competes directly with names like DeepSeek V3 and Qwen 2.5.
What stands out in the report:
📊 On MMLU-Pro, Hermes 4-405B scores above DeepSeek V3 671B (MoE) and close to Qwen 2.5-72B, placing it in elite territory.
🧩 On reasoning-heavy benchmarks like GPQA Diamond, Hermes 4 shows gains over many open-source peers, a sign it handles multi-step logic well.
🧠 Comes with a “Cogito” variant, designed to log its chain-of-thought. Instead of hiding its reasoning, it literally thinks out loud.
🍟 Fun detail: asked for a Lovecraft-style poem about fries, Hermes 4 first drafts a horror “outline” with creepy steps before writing the poem itself. Surreal and transparent.
Why it matters:
1. Few labs give the public raw reasoning traces, Hermes 4 does. That’s gold for researchers, jailbreakers, and alignment studies.
2. It shows @NousResearch is doubling down on big open models while others hide weights.
3. If released on @HuggingFace (which seems imminent), this could be one of the largest openly available models ever.
#Hermes4 #LLM #OpenSourceAI




English
















