Kronos is a self-evolving training pipeline targeting Opus-4.6-parity
on coding benchmarks at $1/M output tokens,
2M context, 150–200 tok/s.
All scripts, datasets, and intermediate checkpoints are public.
€100
seed capital
€5.50
spent so far
Qwen2.5-Coder-7B
Round-1A base model
HumanEval+ / LCB
evaluation benchmarks
Status (May 11 2026)
Round-0 DONE: Qwen2.5-Coder-1.5B + LoRA r=32 on 30K CodeFeedback rows. 625 steps, loss 0.678→0.489, mtok-acc 0.813→0.841. Adapter at jaivial/kronos-round0-qwen15coder-lora.
Round-1A v3 LIVE: Qwen2.5-Coder-7B + attn-only LoRA r=16 on 15K CodeFeedback. Training on Kaggle P100, ~12h wall-clock. Lean config calibrated to fit Kaggle 12h cap after v1/v2 hit cap-killed at 13% trained.
Round-2 wired: distillation (Branch A) from Qwen3-Coder-480B-A35B teacher, OR more SFT (Branch B), OR diagnose (Branch C). post-round.py auto-selects from R1A's HumanEval+ pass@1.
Operating stack: HF Pro (~€5.50/mo prorated) for 480B teacher access. Kaggle P100 free for training. R6 serving stack (vLLM + DFlash speculative decoding) architecture sketched.
Approach
Cheap-first ladder: Kaggle/Colab free GPUs → grant credits → small paid GPU bursts → revenue-funded scale.
Distillation Round 2: Qwen3-Coder-480B teacher fixes ~50% of student failures on LCB-medium (empirically validated cycle 38–39).
RL Round 3: GRPO + binary code-execution reward on the residual systemic failures distillation can't fix.
Speculative decoding Round 6: DFlash block-diffusion drafts for 2–3× throughput at serve time.
Early access
Want to try the model when Round-3 lands? Drop your email.
Paid early-access tier (€10–20/mo with usage cap) opens after the first Opus-4.6-comparable checkpoint.