Carnice V3 27B — written Carnis-V3-27B on the model card — dropped on 24 August 2026, and it is a genuinely different kind of release: a 27.8 billion parameter open-source model built on the Qwen 3.8 27B base, trained specifically for agentic tool use rather than chat, with its creators claiming it can compete with models ten times its size — and the whole thing runs locally on a consumer GPU. No API, no cloud, your machine. I read the actual model card before filming my breakdown, and one detail in the training data changes how you should think about this model — so here is the full picture, not just the headline.
📺 Watch: NEW Carnice V3 27B Just Dropped
🔥 Get the Agent OS as a free bonus: AI Profit Boardroom members get the full Agent OS zip, prompt libraries, daily tutorials and weekly live coaching calls. → Get inside
Regular readers know I test local agent stacks constantly, because local is where the economics of AI automation get silly — no per-token bill, no rate limits, and your business data never leaves the building. Carnice is interesting precisely because it was not trained to be a better chatbot. It was trained to take a task, reason through it, call tools, observe the results and keep going until the job is finished. That is agent behaviour — a completely different thing from a chat window.
What Carnice V3 27B Actually Is
The essentials, straight from the model card I walk through in the video: Carnice V3 is a 27.8 billion parameter model built on the Qwen 3.8 27B base, released under a full Apache 2.0 licence — open, free, and with no restrictions on commercial use. The base model is almost a side detail, though. What matters is the training target: agentic tool use. Where most fine-tunes chase nicer prose or better benchmark chat scores, this one was tuned for the loop that actually does work — task in, reasoning, tool calls, observation, next step, job done.
The bold claim attached to the release is that a model this size can stand in for models around ten times larger on agent work. Treat that like every release-day claim: interesting, and unproven on your workload. It goes through Goldie Bench, my own benchmark, like everything else I cover — same tasks, same prompts, scoreboard decides — and I would run your own version of that discipline before moving anything important onto it.
The Brain and Body System: Carnice V3 27B Plus Hermes Agent
Carnice was built specifically around Hermes Agent, the open-source agent runtime from Nous Research, and the simplest way to understand the pairing is the one I use in the video. The model is the brain. Hermes Agent is the body. The tools — files, browser, terminal, memory, scheduling — are the hands. When all three work together, you stop getting a chatbot and start getting something that can actually do work. That is the real story here, and it is why this release matters more to my audience than another incremental chat model.
If you have not set up the body yet, my guide to installing Hermes with a free local model covers the runtime side, and my walkthrough of running Hermes on local Ollama models shows the day-to-day workflow once it is wired up. Carnice slots into that same architecture as the brain — one more reason the model-agnostic setup work pays off every time something new drops.
Inside AI Profit Boardroom we have a full 30-day local agent road map — installing Hermes Agent, connecting models like Carnice, and building automation workflows that run real business tasks — plus four live coaching calls every week → Get the local agent road map
📺 Watch: Qwen 3.8 27B is NOW on Ollama... This is CRAZY
Hardware: What You Need to Run Carnice V3 27B Locally
This is what makes Carnice interesting for people who are not running data centres. The full BF16 checkpoint is around 55 GB, but the model also ships in GGUF format, which means quantised versions run on consumer hardware. The sizes break down like this: Q4_K_M at 15.66 GB is the smallest recommended build; Q5_K_M at 18.19 GB is the recommended balance; Q6_K at 20.89 GB buys higher quality; and Q8_0 at 27 GB is maximum fidelity. In practice, if you have an RTX 3090 with 24 GB of VRAM, Q5_K_M fits and you run the whole agent locally.
Local matters for more than cost. Your data never leaves your machine — and for anyone running workflows over private lead lists, member data or business strategy documents, that is not a nice-to-have, it is the whole argument. If you are weighing Carnice against the other options at this size, my rankings of the best local models for a Hermes agent and the best Hermes agent models overall give you the field it has to beat — this page covers the new release itself, those pages tell you what I would actually pick today.
The Training Details That Genuinely Surprised Me
Now the part I flagged at the top. The training run behind Carnice, per the model card: a rank-64 RS-LoRA fine-tune on the Qwen 3.8 27B base — meaning around 466.9 million parameters were trained across 496 target modules and merged back into the full model, rather than retraining all 27 billion from scratch. The run was 24 steps in total, with step 12 selected as the final checkpoint because that is where validation peaked. The whole thing ran on a single 96 GB NVIDIA GPU and took about 44 minutes.
The data is the eyebrow-raiser: eight private trajectories across six agent task families — 162,798 supervised tokens in total. For perspective, big model training runs use hundreds of billions of tokens; this fine-tune used 162 thousand. Read that both ways. Charitably, it says the base model plus a small amount of extremely targeted agent data goes a long way, which would be great news for anyone building specialised local agents. Sceptically, it says the "competes with models ten times its size" claim rests on a very small, private training set, and you should verify behaviour on your own tasks before trusting it with anything that matters. That is exactly why I say in the video: worth testing carefully, worth watching closely, clear eyes about what it can and cannot do yet.
📺 Watch: How to Run Hermes Agent for FREE
Two Workflows I Would Run With Carnice V3 27B
Ranked by immediate business value, here is where I would point it first — both are workflows I demonstrate in the video, running through Hermes with Carnice as the brain.
- Member and lead intelligence agent. Give the agent a file of leads who signed up for a free resource or showed interest, and have it segment them by buying intent — high, medium, low — explain its reasoning for each group, build a prioritised outreach plan, and draft a short personalised message per segment. That is hours of manual work done in a single agent run, and because it is local, the lead list never leaves your machine.
- Weekly content machine. Drop in a folder of the week's AI research and updates, and have the agent produce a weekly content plan: LinkedIn posts, email subject lines with preview copy, video concepts with hooks and outlines — each saved to its own labelled file. You review, edit, publish. One prompt, a full week of content scaffolding.
Both of these are straight out of the playbook I run through the Agent OS, the agent operating system I built and test daily in my own business — the model in the brain slot changes, the workflows stay. And if budget is the constraint rather than hardware, my guide to the best free AI models for a Hermes agent covers the zero-cost end of the same decision.
Carnice V3 27B: Key Specs
| Spec | Detail, per the model card |
|---|---|
| Model | Carnis-V3-27B — 27.8 billion parameters, built on the Qwen 3.8 27B base |
| Licence | Apache 2.0 — open, free for commercial use |
| Training target | Agentic tool use: reason, call tools, observe, continue to completion |
| Built for | Hermes Agent, the open-source agent runtime from Nous Research |
| Full checkpoint | Around 55 GB in BF16 |
| GGUF sizes | Q4_K_M 15.66 GB · Q5_K_M 18.19 GB (recommended) · Q6_K 20.89 GB · Q8_0 27 GB |
| Consumer hardware | Q5_K_M fits an RTX 3090 with 24 GB VRAM |
| Fine-tune method | Rank-64 RS-LoRA: ~466.9M trained parameters across 496 modules, merged back |
| Training run | 24 steps, checkpoint from step 12; single 96 GB NVIDIA GPU; about 44 minutes |
| Training data | 8 private trajectories, 6 agent task families, 162,798 supervised tokens |
Carnice V3 27B FAQs
What is Carnice V3 27B?
It is a 27.8 billion parameter open-source model — written Carnis-V3-27B on the model card — built on the Qwen 3.8 27B base and trained specifically for agentic tool use, released under Apache 2.0 on 24 August 2026. It is designed to act as the brain of an agent stack rather than a chat model.
Can I run Carnice V3 27B on a consumer GPU?
Yes — that is the headline feature. It ships in GGUF format, and the recommended Q5_K_M build at 18.19 GB fits a 24 GB card like an RTX 3090, so the full agent runs locally with your data staying on your machine.
Does Carnice V3 27B really compete with models ten times its size?
That is the creators' claim, and the model card shows it was fine-tuned on a very small, private dataset — 162,798 supervised tokens. Promising, but unproven until you test it on your own workload. Run your own benchmark tasks before trusting it with anything important.
What do I need besides the model?
An agent runtime. Carnice was built around Hermes Agent by Nous Research — the model is the brain, Hermes is the body, and the tools are the hands. Install the runtime, load the model, and the agent loop does the rest.
Is Carnice V3 27B free for commercial use?
Yes — it is released under a full Apache 2.0 licence with no restrictions on commercial use, per the model card.
Verdict: Early, Real, and Worth Testing With Clear Eyes
Carnice V3 27B is an early but real step towards something my audience has wanted for a while: a genuinely capable agent brain that runs on hardware you already own, with a licence that lets you build a business on it. The hardware story is excellent, the Hermes pairing is exactly right, and the training details are both the most impressive and the most cautionary part of the release — a 44-minute fine-tune on 162 thousand tokens making 10x claims deserves testing, not blind trust. Point it at recoverable workflows this week, watch what it does, and let your own results decide.
The moment tools like Carnice drop, we are inside AI Profit Boardroom building workflows around them — step-by-step setup tutorials, four live coaching calls a week, and members already running local Hermes agent stacks on real business tasks → Test Carnice with us inside AIPB
