Results, models, and notes from the Belna team.
Our flagship agent pairs open Kimi K3 weights with a privacy-first harness — and holds its own against the biggest frontier models on agentic benchmarks. Here's what it can do, how it scores, and how the safety layer works.
Read article →Temporal Associative Memory adds a learned recurrent world-state alongside attention. Trained on 2B tokens against a matched Transformer, it wins on loss at every scale — with full training code so you can replicate it.
Read article →Pin every word to meaning axes like hot–cold and true–false, and a tiny model starts to reason: perplexity 80.89 vs 111.45 against its baseline, 35/35 on negation, embeddings where opposites are genuinely opposite.
Read article →Sparse mixture-of-experts meets persistent recurrent state. A completely independent design, the same 2B-token protocol — and the same result: lower loss than the Transformer. Safety bounds included, training code included.
Read article →A single Transformer layer looped four times around a memory database with a symbolic router. At 11.3M parameters it halves its baseline's perplexity (63.89 vs 111.45) — and cracks all 35 negation probes with zero help at inference.
Read article →Everything running in production or under active investigation — switch tabs to explore.
The personal agent that does the work — 300 clicks across 4 apps replaced by one sentence. Open Kimi K3 foundation, 1M-token context, Smart and Fast modes, wrapped in the Belna harness with approvals, a sealed vault, and a visible trace.
Our 23.5M-parameter model that pins words to meaning: topology coordinates learned jointly with language modeling on WikiText-2. Lower perplexity than its baseline (80.89 vs 111.45), a 0.957 meaning correlation, and 35/35 on negation with axis guidance.
One Transformer layer, looped four times around an explicit 10-slot memory with a symbolic negation router. At 11.3M params it beats its matched baseline on perplexity (63.89 vs 111.45) — and solves negation 35/35 with no help at inference.
Our 101.8M-parameter hybrid: reduced-width attention running in parallel with a recurrent world-state, mixed by a learned gate. Lower loss and perplexity than a matched Transformer at 25M, 50M, and 100M — plus a real memory advantage at 256-token contexts.
An independent cross-check: 8-expert sparse MoE with persistent recurrent state and full attention only every 6th layer. Same data, same hardware, same seed discipline — and it also beats the Transformer on loss, with auditable, bounded compute.
Every paper ships its full training code inline — no hidden repositories. Replicate, remix, improve.