Models

Clean spherical KAN checkpoint at step 72,500 after 4.751B ordered byte tokens (validation BPB 1.1512). No chat, instruction, EOS, or public-eval fine-tuning; Q:/A: is endpoint framing only.

Public-eval-contaminated benchmark specialist derived from the 128.26M-parameter Harmonic GPT spherical SpectralKAN/Lohe byte model. Only final normalization and the byte readout were post-trained; all oscillator dynamics remain frozen. Trained directly on the public eval prompt snapshot, so scores are training-set measurements rather than generalization evidence. Audit: 18/28 exact greedy, 28/28 null-stopped, 92.07% teacher-forced byte accuracy.

Versioned negative-result checkpoint from Alpha's V11 all-token bridge. It was trained over 10,862 GPT-5.4-generated, GPT-5.5-reviewed synthetic conversations with Symbiogenesis disabled. On the fixed 615-conversation diagnostic it produced 615 nonempty EOS-terminated replies but 12 detected loops. In a 100-case reference-blinded GPT-5.5 review it scored 15 PASS / 11 BORDERLINE / 74 FAIL; the V8 reference ranked first, and the reviewer selected NONE. Quality gate: FAIL. Standard Transformers export: huggingface.co/ajaxdavis/alpha-chat-v11-m300-experimental at revision 29f0372fb94c1d249421daca50c3fbd263dc1309.

Experimental 127.6M-parameter raw-byte chat Transformer with public weights and ZeroGPU OpenAI-compatible inference.

98.7M-parameter decoder-only model grown on a single RTX 4090 with a bandit-controlled curriculum over open data, as part of a permacomputer project. Deliberately neither GPT-2 nor LLaMA: parameter-free RMSNorm, learned absolute positions, a norm on the embedding sum, no final norm, ReLU MLP, no biases, untied LM head. Tokenizer is Harris (1955) successor-variety segmentation over a 256-byte fallback, 8192 learned morpheme segments, discovered from the distributional properties of text with no language-specific rules. Trained on megachat-v8 (chat, oasst, dolly, gutenberg, IRC, dictionary and more) through a phase-based UCB1 bandit that reweights sources by their own loss trajectory. Weights, config and reference implementation are open: huggingface.co/russellbal/ANDREA-120M (AGPL-3.0). Served here from a ZeroGPU Space.

Alpha is a 57,688,576-parameter Llama-form conversational model trained entirely through a from-scratch TypeScript tensor library, tape autograd, hand-generated SPIR-V and a hand-written Helios Vulkan backend — every training FLOP on the project's own code, no PyTorch/CUDA in the training path (PyTorch was used only to verify the export). 16 layers, hidden width 512, 8 heads, SwiGLU 1408, 512-token context, RoPE + RMSNorm, tied embeddings, 12,288-token byte-level BPE. This is the corrective "chat repair" checkpoint (step 1,200), selected by deterministic free generation rather than by validation loss: step 2,000 had the better teacher-forced loss (1.6814) but failed to stop on two prompts and repeated more. It replaces a terminal SFT artifact that returned an empty string on 92 of 100 frozen prompts. Honest status: structurally chatty, semantically immature. On the 48-case repair suite it produced 48/48 nonempty, EOS-terminated replies with no role leaks — but 5 still hit the repetition-loop threshold, and it is shallow, weak on idiom and instruction-following, and frequently factually wrong. It self-reports quality_gate=FAIL. Treat it as an inspectable research artifact, not a dependable assistant. Served from the project's own inference server (Apache-2.0; weights at huggingface.co/ajaxdavis/alpha-60m-chat).