Open Source Radar — November 2: 290B models on a desktop GPU
Today's board is about putting big models on hardware you already own — one serving engine for gaming PCs, one foundation model for stock charts, and one security operations center that runs without a single API key.
FreeToken (Python, 15,691 stars, Apache-2.0) is an inference engine built for the opposite of a datacenter: it runs frontier-scale mixture-of-experts models — 290B parameters and up — on a gaming PC by treating your GPU, CPU and system memory as one elastic pool, according to its repo. The project comes out of UC Berkeley's Sky Computing lab, with Ion Stoica, Matei Zaharia, Song Han and Kurt Keutzer among the authors of its paper, which frames the trick as bandwidth-adaptive execution — whichever hardware can feed an expert fastest gets the job. It already lists DeepSeek-V4-Flash, Qwen3.6-35B-A3B and GLM-5.2 as supported, speaks the OpenAI and Anthropic APIs so coding agents plug straight in, and runs on RTX 30, 40 and 50 cards. The honest caveat is speed: local MoE serving on one consumer card is usable for agent loops, not for serving a team, and the repo's own pitch is about interactivity, not throughput.
Kronos (Python, 45,697 stars, MIT) is the first open-source foundation model trained on financial candlesticks rather than text, and it jumped back onto daily trending today despite a last code push in April. The design is clean: a tokenizer turns OHLCV bars from more than 45 global exchanges into discrete tokens, then a decoder-only transformer predicts the next ones, so one model handles forecasting, volatility and synthetic-data tasks that used to need a bespoke model each. The released checkpoints are small — 4.1M to 499M parameters — and the paper was accepted at AAAI 2026. Two caveats worth knowing before you point it at a brokerage account: the 499M "large" model is deliberately not open, and a candlestick model learns chart patterns, not fundamentals or news.
AiSOC (Python, 6,503 stars, MIT) is a self-hostable security operations center where an LLM agent does first-pass triage: it fuses alerts from your existing tools, groups what fires into incidents, investigates each against MITRE ATT&CK, and records every prompt and tool call in a replayable ledger before a human approves any response, per the project's documentation. The detail that makes it worth a look is the no-keys default — it ships a local model, the CISA KEV threat-intel catalog, and a rule library of 7,023 rules with 2,538 marked executable, so a solo admin can stand it up on one host with 8 GB of Docker memory and no cloud contract. It also exposes an MCP server for Claude, Cursor and Continue. Treat the AI verdict as a queue-sorter, not an analyst: the repo is candid that the agent proposes and a person decides, which is the only sane posture for automated response.
Worth watching this week: whether FreeToken's edge serving gets benchmarked on hardware people actually own — the 290B claim is a paper result until someone posts tokens per second from a 24 GB card.
If an LLM agent flags an alert at 3am, who signs off on the block? Write to us via our contact page.
