Agents-A1 just dropped.An agent brain that runs free on your Mac.
InternScience's new open model is built for one thing: agent work — long searches, tool calling, engineering. It's a 35B mixture-of-experts that only lights up ~3B per token, so it runs at 95 tokens a second on my Mac, for free. I put it through all 45 GoldieBench builds the day it dropped. Real scores, real builds, honest verdict below.
The official links. Read it, run it yourself.
"Optimized for long-horizon search, engineering, scientific research, instruction following, and tool calling" — with claimed state-of-the-art results on Seal-0 (56.36), IFBench (80.61) and BrowseComp in its class (75.51).
— InternScience Agents-A1 project page, July 2026
The local agent brain has always had a catch.
Before
Every local model I tested made me pick one: fast or capable.
The small ones flew — and fumbled anything longer than a sentence of instructions.
The big ones could think — and crawled so slowly an agent loop took all afternoon.
So my agent loops kept defaulting back to cloud models, with a meter running.
Then a lab shipped a model built for exactly this: a big brain that only wakes the part it needs.
After
Agents-A1 is 35B of knowledge that runs at 95 tokens a second on my Mac.
It's tuned for the exact shape of agent work — long instructions, tools, multi-step jobs.
And it's free, offline, and on the leaderboard the day it dropped.
Fast AND capable, on the machine you already own. That's new.
Why listen to me on this.
I run an AI agency with 70+ people where AI handles about 80% of the ops, and I bench every model that matters on the same 45 one-shot builds at goldiebench.com.
No invented quotes. The wins are real and written by the members themselves — agency owners, ecom founders, creators, solo operators across 38 countries. Read them in their own words.
Read the 158-page wins doc →Commit to running one new model this week.
Agents-A1 is two commands away from running free on your Mac. The exact commands are below.
Here's the deal I want to make with you.
This week, pull ONE new local model and give it a real job — a draft, a summary, an agent loop. Just one.
Because the people who test each new model the week it drops are always sitting on the best free brain available — while everyone else pays for last season's default.
Be one of them.
Commit to the transition. One new local model, one real job, this week.
A 35B brain that only wakes 3B at a time.
Agents-A1 is a mixture-of-experts model. All 35 billion parameters hold the knowledge, but each token only activates about 3 billion of them — so you get big-model answers at small-model speed. Then InternScience tuned it in three stages (supervised fine-tuning, domain teachers, multi-teacher distillation) specifically for agent work: long-horizon search, engineering, instruction following, tool calling. 256K context.
The result on my M4 Max: 95 tokens a second, fully local, from the official 21GB Q4 build. That's the fastest local model on my machine — by a lot:
45 one-shot builds. Here's where it landed.
Same GoldieBench gauntlet every model gets: one prompt per task, no retries, no help — then a real rendered screenshot goes to an Opus 4.8 vision judge that scores 0–10 on the same bar as the frontier models. Agents-A1 finished at 4.83 average across all 45 tasks — the #2 local model on the board, with 3 local golds:
The honest read: Qwable (a dedicated 27B coding model) still owns the local crown for one-shot builds. But Agents-A1 isn't a coding tune — it's an agent tune — and it still beat every other local model on the board while running roughly 4–5× faster than Qwable. Ten of its builds scored 7 or higher. The weak tail is the brutal graphics tasks (path tracers, shaders) that flatten every local model.
What it actually built. Click in and play.
Every one of these is a raw one-shot build from Agents-A1, running locally — exactly as the judge saw it. Click any card to open the live build.
That landing page at the top scored 8.1 — frontier-tier for that task — from a free model on a laptop. And the Dragon Realm 6.3 is the strongest local one-shot I've seen on the flagship task outside the dedicated coding models.
Two commands. Free. Offline.
ollama pull hf.co/InternScience/Agents-A1-Q4_K_M-GGUF:Q4_K_M
ollama cp hf.co/InternScience/Agents-A1-Q4_K_M-GGUF:Q4_K_M agents-a1That's the official 21GB Q4 build — it wants ~22GB of RAM, so it's happiest with 32GB+. It's a thinking-family model, so give it room (or set think: false for straight answers). From there it's ollama run agents-a1 — or wire it into your agent stack as the local brain.
Wire Agents-A1 into the Agent OS.
A fast, free agent-tuned brain is exactly what local agent loops have been waiting for. The Agent OS is where it plugs in next to everything else — one dashboard, one memory.
You're not buying a tool. You're getting the whole operating system I run a seven-figure business on.
Get the Agent OS →No — models like Agents-A1 are exactly why not. The Agent OS runs the everyday 90% on a free local model on your own machine — and this one is a 35B agent brain at 95 tok/s, at $0. Free APIs slot in for more, and for the frontier work it drives the CLIs you already pay for (your Claude subscription already includes the Claude CLI — the Agent OS plugs straight into it, so you're not paying twice).
It's a layer on top of what you already own, not a new meter. And inside the Boardroom there are full token-optimisation tutorials so you cut usage to the bone.
How people pick a local model. And how the bench does it.
- Read the benchmark table in the announcement
- Assume "SOTA" means it fits your use case
- Never actually run it on your own machine
- Pick by hype, keep by inertia
- No idea how it compares on YOUR tasks
- Pull it the day it drops, verify it generates
- Same 45 one-shot builds every model gets
- Real rendered screenshots, vision-judged 0–10
- Every demo playable on the public board
- Honest verdict: where it wins, where it doesn't
What's holding you back.
Wrong: "A 35B model can't run on a normal Mac."
Right: The MoE design changes the math — 35B of knowledge, ~3B active per token. The official Q4 build is 21GB and runs at 95 tok/s on a 36GB Mac. Verified, not claimed.
Wrong: "Local models are toys for hobbyists."
Right: This one one-shotted an 8.1-scoring landing page and a playable racing game, and it's tuned for the tool-calling agent work businesses actually run. Free and offline is a feature, not a compromise.
Wrong: "I'll wait for the model wars to settle."
Right: They won't settle — a better free brain lands most weeks. The edge is a system that tests and swaps models in a day, so you're always running the current best.
158 pages of members already running this stack — real businesses, real wins, in their own words.
Read the 158-page wins doc →The best free brain changes weekly. Be ready for it.
Agents-A1 is this week's drop — an agent-tuned 35B running free at 95 tok/s on a laptop. Next week there'll be another. The winners aren't the ones who picked the perfect model; they're the ones set up to test and swap in a day.
Inside the AI Profit Boardroom you get the full Agent OS — the local engine ready for any Ollama model, Apollo the voice copilot, Agent Kanban, Free Claude Code, the AI Mastermind, memory that knows your business, every CLI you already pay for in one dashboard, a 30-day roadmap, daily tutorials, coaching calls, and 3,600+ founders across 38 countries building alongside you. Every new model — like this one — gets benched and wired in the week it lands.
It's the operating system I run a seven-figure business on. You get the whole thing.
Get the Agent OS →Pull it tonight, give it one real job, and check the live scores at goldiebench.com/local-models. I'll see you in the next one.
