Full-stack developer & digital product builder. I build useful software and write about AI news, tools and experiments.
Italia
Joined February 2021
- Tweets108
- Following373
- Followers22
- Likes268
Open-sourced Vastloom: a C++20 runtime that runs a 175 GiB MoE checkpoint end-to-end on a 31 GB RAM desktop, streaming weights from NVMe.
Greedy parity with llama.cpp, honestly ~10x slower. Every claim measured.
github.com/TheStreamCode/vas…
Michael retweeted
Meet Mistral Large 4, aka Le Chonk.
• 1T parameters, natively multimodal. 49B active.
It is the best open weights model from US or Europe on aggregated benchmarks.
• State-of-the-art on critical workloads, including cyber defense, manufacturing and finance and it surpasses closed frontier models on visual grounding.
• Forged in Europe end-to-end and is deployable from Europe via our own Mistral Cloud infrastructure.
• Available to all via API today. Working with cybersecurity partners privately.
Open weights release end of October.
🤖 Made with AI
chutes-provider is now in the official Hermes Agent catalog.
Chutes.ai models via API key, OpenAI-compatible endpoint, 13/13 checks passing.
hermes plugins install chutes-provider
github.com/NousResearch/ hermes-agent/pull/126683
openclaw-muse-code v0.2.0: live model discovery (60s TTL, static fallback) + security hardening — minimal credential cache, ClawHub scan clean. Muse Spark on your subscription, no API key.
github.com/TheStreamCode/ope…
Ship Muse Spark on your Muse Code subscription, no API key, no extra CLI. Device-code login, Responses API, verified end-to-end. For Hermes and opencode (npm).
Upstream PRs open:
- github.com/TheStreamCode/her…
- github.com/TheStreamCode/ope…
Hermes Agent's Meta provider only takes API keys, so Muse Code subscribers get metered twice. I built a provider plugin that bills Muse Spark to the subscription instead.
hermes plugins install TheStreamCode/hermes-muse-code
Repo + catalog PR upstream: github.com/TheStreamCode/her…
Some digital experiences are built to inform. Others are built to be felt.
Valmont is a new web concept I designed and developed around motion, rhythm and a strong visual identity.
Explore the experience: mikesoft.it/concept/valmont.…
#WebDesign
T3 Code is an open-source interface for Codex, Claude Code and OpenCode. The interesting idea is not another chat UI—it’s treating multiple coding agents as interchangeable backends for the same workflow.
Apple's new Mac Studio M5 Ultra configures to 256GB of unified memory at 1.2TB/s. DeepSeek's open V4 Flash 0731 (284B params, 13B active) has a lossless 8-bit GGUF at 162GB. Do the math: a top coding model, fully offline. No API key, no subscription.
Model leaderboards are useful only when you know what they measure. Intelligence, coding, latency and price often point to different winners. Artificial Analysis is a strong starting point—not a substitute for testing models on real workloads.
A "stealth" model appeared on OpenRouter last Thursday with no name behind it. Ox Alpha: free preview, 1M context, built for coding and long agent runs. Anonymous provider, speculation everywhere. Even Stripe's CEO calls it very impressive.
openrouter.ai/stealth/ox-alp…
DwarfStar is not another generic local LLM runner. It is a C inference engine built specifically for DeepSeek V4 Flash, with OpenAI, Anthropic and Responses API support. The trade-off: serious hardware requirements and a model-specific stack.
Codex isn’t limited to OpenAI models. Its OSS mode supports local providers such as Ollama and LM Studio, while custom providers can use OpenAI-compatible endpoints. The coding harness and the model are separate choices—and that reduces lock-in.