I just wrapped up the first research module on local agentic AI and I have a lot to say. Before I start publishing my work on agents, I wanted to drop this post as a warmup (and a warning). We are progressing rapidly with local AI. As of 2026, we fast-tracked our way via model releases that stepped up the game to the critical event 2 months into 2026 – the release of Qwen 3.5 27B (February 24th). As time progresses, we will look back at this moment as a major catalyst event in local AI.

• 2024: The year local AI blew up (llama 3 family + Chinese MoE wave)
• 2025: The year of Reasoning and MoEs (DeepSeek R1 and Qwen3-Coder-480B-A35B)
• 2026: The year of Agents (Qwen 27B variants, Google Gemma 4 variants)

2026 is the Year of the Horse. The Chinese Zodiac (生肖) contains 12 animals with 5 elements: Wood, Fire, Earth, Metal and Water. For Lunar New Year purposes, the Year of the Horse began on February 17, 2026 and it will run until February 4, 2027. The way the Chinese Lunar Calendar works is that each animal repeats every 12 years and this year is not just the Year of the Horse – it is the Year of the Fire Horse. The animal + element combination only repeats every 60 years – the last Fire Horse was 1966; the next will be 2086.

In traditional meaning:

Horse = energy, movement, independence, confidence, speed, ambition, freedom.

Fire = passion, intensity, visibility, courage, action, heat, emotion.

The Fire Horse is often interpreted as a year of bold movement, big changes, high energy, independence, risk-taking, and momentum. It is considered a powerful, fast-moving year where people may be pushed to act, make decisions, start things, or break away from old limits. Fire Horse energy can also mean impulsiveness, burnout, conflict, impatience, ego, or doing too much too fast. The traditional advice is basically – “use the courage and momentum, but don’t be reckless.”

TL;DR: Fire Horse = “move forward boldly, but don’t let passion turn into chaos.” Symbolically, it is considered one of the most intense and dynamic zodiac years in the Chinese Lunar Calendar.

As I write this post, we have already jumped from Qwen 3.5 27B to Qwen 3.6 27B – an improvement over an already amazing local model. I have not yet tested this model, as I am still blown away by Qwen 3.5 27B. This was the first open-sourced AI model where you could run serious agentic tasks via the Hermes Agent on consumer-grade hardware – meaning, RTX 3090s. I can’t speak for CPUs, but I assume that running this thing on CPU inference is just as amazing (a possible research area when I finish my new GPU cluster build – the v2). We are most certainly in the Year of the Fire Horse, and we are just getting started.

Almost 3 months into the Year of the Fire Horse – and we still have another 9 months to go (and we’re nearly halfway through 2026 in general). The average person doesn’t realize how important the Qwen release was and is. This is really the first time we have a really strong model that can: think, execute tool calls, do agentic work (not just agentic coding) all on consumer GPUs. The key here is consumer GPUs. It’s one thing to run Kimi K2 (a 1 trillion parameter MoE) on four RTX PRO 6000s at aggressive quantization – but to get this level of quality in a model on two RTX 3090s – that is beyond impressive, there has been a full Overton window shift.

I wanted to write this prelude before I start publishing a lot of agentic AI research for a few reasons.

1. I am going to publish a lot of work on the Hermes Agent by NousResearch and local AI models. The Year of the Fire Horse is also the Year of the Agents and I, too, have fallen into the hype trap.
2. There has been a major Overton window shift with Qwen 3.5 27B. This is a focal point that I am strongly paying attention to – not just from an AI researcher POV – but just a general technology shift. This is going to affect many things at once, such as the gaming GPU market, memory prices, etc. The models are getting smaller and better. One can make a strong argument that distillation attacks are getting better etc., but regardless, this is a serious thing to be paying attention to with the Chinese models.
3. This is a major warning. People need to start waking up now. Local AI is closing the gap on frontier models.
   • with Moore’s Law effectively coming to an end on the current way we create and use GPUs, the Scaling Laws will be tested (see Jensen’s comment on this from the infamous Dwarkesh podcast)
   • frontier labs such as Anthropic pulling 3rd party support for Claude Code is a major signaling event that local AI development has shaken up the big labs
   • China is winning the open source model front
   • white-collar jobs are going to get hammered within the next 18-24 months (hopefully I am wrong)
4. I will be posting 2-3 research modules on local agentic AI – specifically with the Hermes Agent. I have just completed the first module, and it’s safe to say that I am blown away. Dario is absolutely sweating, no doubt, as the IPO is on the horizon, and their breadwinner, Claude Code, may not be the best product on the market for much longer. This 2-3 research series will focus on consumer-grade GPUs running local agents. I am trying to structure it in a way where:
   • I am benchmarking different local models
   • Pushing the limits of consumer-grade hardware (RTX 3090s)
   • Creating guidelines for anyone out there who wants to run agents locally
   • Supporting local AI
   • Attempting to replace some of my agentic AI usage with Claude Code by doing this

The Year of the Fire Horse is shaping up to be exactly what it’s supposed to be. Now is the best time to buy a GPU and start working on local AI. Put your UPS/FedEx hat on and start delivering everyone, because things are about to move very fast and powerfully.

This is part of Project Arrival, my journey from Senior Data Scientist to AI research engineer. I’m documenting the build, the research, and the path to joining a top AI lab. Follow along at blog.zacharycangemi.com or at my GitHub.