Maxim 1.0 — The Autopsy
Alflora Research · 01 Aug 2026
Tl;dr
- Maxim 1.0: a 240M-parameter coding model, trained on a single Arc A770.
- It survived three full architecture migrations: Mixture of Experts, then LLaMA, then GPT.
- The vision attempt hit the compute wall and the project stopped — cause of death was a budget, not a bug.
- Every lesson from the corpse went directly into Maxim 2.0.
The patient
Maxim 1.0 was the first serious attempt at training an own model: 240 million parameters, aimed at coding question-answering, trained on the one GPU this rack has — the Arc A770 16GB. By big-lab standards that's a rounding error. By bedroom-server standards it was an ambitious organism, and for a while it was very much alive.
It answered questions. It learned syntax. It hallucinated with confidence, as is tradition. Most importantly, it was ours — every weight trained on hardware we could touch, every checkpoint restorable without asking anyone's permission.
Three architectures
The project outgrew its own skeleton three times. It started as a Mixture of Experts, because routing sounds efficient on paper and looks brilliant in diagrams. Then the training instabilities and the sheer plumbing cost of MoE on modest hardware forced a migration to a LLaMA-style architecture — simpler, denser, honest. When tooling and ecosystem gravity kept pulling toward the GPT shape, it migrated again.
Each migration meant retokenizing, reshaping, retraining from earlier checkpoints than anyone would like. Each one also taught exactly why the field converged on these designs — not from papers, but from watching the loss curve misbehave in person.
The vision attempt
The ambition that ended it: give Maxim eyes. A vision encoder, multimodal training, screenshots in, answers out. The plan was sound; the arithmetic was not. Vision training multiplies sequence lengths and memory pressure, and 16GB of VRAM with a part-time GPU is a wall, not a hurdle. Training slowed from days to geological time, and the honest conclusion arrived: this organism cannot grow on this power supply.
So the project was halted — deliberately, documented, with checkpoints kept. Not a crash. A controlled landing into the ground.
Cause of death
Write it plainly: not enough POWAHHHH. The ideas weren't the failure; the ideas were fine, and three architectures of evidence say so. The failure was that sustained training compute is a budget, and a bedroom has a ceiling. Compute is a budget before it is an architecture — the model didn't die because it was wrong, it died because feeding it further cost more power than owned hardware could supply.
What the corpse taught
Maxim 2.0 exists because 1.0's autopsy was thorough. The 2026 revival inverts the equation: rent the brain through OpenRouter — currently Kimi K3 — and own the body: accounts, history, streaming, serving, all on this rack at maxim.alflora.app. Borrowed power where borrowing is cheap, owned power where ownership matters. The evaluation habit stayed too: models are swapped when testing says so, not when marketing says so.
A dead model that teaches you where the walls are is not a failure. It's tuition.