In mid‑October 2026 I posted a mix of product promotions, AI model announcements, and personal workflow notes on X. The highlights include adopting the free e‑signature service mndok, hitting my Claude usage limit, and evaluating a slew of newly announced open‑source models for future custom harnesses.
mndok is a 100% free, unlimited document signing platform with no hidden paywalls, no credit cards, and zero limits. It supports unlimited envelopes & signers, sequential/parallel workflows, audit trails & certificates, and full API access.
This is first time I reset my limit in Claude, and now I’ve exhausted my limits till next Wednesday apparently.
Psychology too is biology based, which is chemistry based, which is physics based, which is finally math based.
- NVIDIA Nemotron 3.5 Super VL: 120 B total parameters, 12 B active.
Decisions: I will adopt mndok for unlimited signing workflows, purchase Claude tokens after the quota expires, and build a custom harness that containers open‑source models (Kimi, GLM, Deepseek, Qwen, Minimax) to replace Claude and Codex in my daily AI pipeline.
- Qwen‑Image‑2.1‑Turbo: 7 B visual generation model, 8 denoising steps, 2K image quality.
- Iris‑3B: 3 B pixel‑space T2I model, no VAE, open weights.
- Underdog Saluki 27B: ~7× smaller than full‑precision, retains 96 % benchmark performance, runs under 8 GB, Apache 2.0.
- Claude Haiku 5.5: ~75 % cheaper to run than Claude Haiku 4.5.
- EmbeddingGemma 2: 740 M parameters, 8K context window (4× text‑only).
- Griffin (video Turing test): 48 % human‑perceived pass rate vs <3 % prior.
- Phonon‑2: 164 MB download, more accurate than Whisper large (10× larger), transcribes 1 h audio in 20 s on a MacBook Air.