Stuck at the cutoff
Anything after training is invisible. The model can’t learn that a fact has changed.
A search engine for AI, not humans
Vium turns the live internet — and real-time feeds like match scores and stock prices — into verified, calibrated knowledge a model can trust. So a small 7B model can reason like a 70B. No retraining. No fine-tuning.
“EU AI Act enforcement timeline?”
Cite [s#] per fact. If disputed or absent, say so — don’t guess.
The problem
A model conflates reasoning with knowledge. Storing facts in parameters fails three ways — worst of all in the small models that fit on local hardware.
Anything after training is invisible. The model can’t learn that a fact has changed.
Rare facts get compressed away — and a smaller parameter budget makes it dramatically worse.
It sounds equally sure whether it knows or not. It rarely knows what it doesn’t know.
What Vium is
One layer behind which Vium connects to paid third-party APIs — search, live-data feeds, data vendors — plus its own crawling. It verifies, calibrates, and distills everything into a single clean package.
A wrapper that detects when a model needs knowledge, fetches the package, and injects it at inference time. A 7B starts performing like a 70B on current, factual tasks.
Don’t make the model bigger. Make its knowledge better.
How it works
The Vium runtime runs a four-step loop around your local model — and only calls out when the model actually needs to.
“What’s the EU AI Act fine cap, and when do the bans start?”
⚠ Confidently wrong — outdated figure, wrong date, invented detail. No way to tell.
“What’s the EU AI Act fine cap, and when do the bans start?”
Prohibited-AI provisions apply from 2 Feb 2025 [s1·s2·s4]. Fines reach up to 7% of global turnover, though sources differ [disputed: some cite 6%]. Per-country SME caps: not established in current sources.
✓ Current, cited, and honest about what’s disputed or unknown.
Decide if the turn exceeds the model’s reliable knowledge — and protect spend.
Pull the right-tier package from the Vium backend.
Render it into context under a citation contract.
The local model reasons over verified evidence.
Adapters for llama.cpp, Ollama, and vLLM. Also reachable as an MCP server, a REST API, and a Python SDK.
Examples
Left, the model on its own. Right, the same model with Vium — current, cited, and honest about what it can’t confirm.
“What’s the current stable version of Python?”
“The latest stable release is Python 3.11.”
⚠ Outdated — answered from frozen training data.“Python 3.14 is the current stable series, released Oct 2025 [s1·s2].”
✓ Current and cited — corroborated.“Who won Wimbledon 2026?”
“Carlos Alcaraz won the 2026 men’s singles final.”
⚠ Invented — the tournament hasn’t happened yet.“Wimbledon 2026 starts 29 June — no result exists yet [unknown].”
✓ Honest about what isn’t known.“How much of global electricity will data centers use by 2030?”
“Data centers will use 8% of global electricity by 2030.”
⚠ One number, stated as settled fact.“Estimates range ~3–8% by 2030, depending on source and method [disputed].”
✓ Surfaces the disagreement instead of hiding it.Live data
Vium also bridges real-time feeds. Your model can answer about a live match score or a moving stock price — values it could never have memorized — and reason over them on the spot. Every live fact arrives stamped with its source and the moment it was read.
“Is Tesla up or down today, and by how much?”
“I don’t have access to real-time market data, so I can’t tell you today’s price.”
⚠ A frozen model simply can’t answer.“TSLA is $337.10, up 2.4% on the day [live · market-feed · 14:32 UTC] — so, up.”
✓ Live value, timestamped — and the model reasons over it.“Who’s winning the India vs Australia match?”
“I’m not able to follow live matches or current scores.”
⚠ No live feed, no answer.“India 248/4 (38.2 ov), ahead by 56 runs [live · sports-feed · 09:14 UTC].”
✓ Current state of play, pulled the instant it’s asked.Integration
No retraining. No fine-tuning. No pipeline to wire up. Wrap any local model and set vium=True — the runtime handles detecting when knowledge is needed, fetching it, and injecting it.
# wrap any local model — llama.cpp · Ollama · vLLM from vium import ViumModel llm = ViumModel("llama3:8b", vium=True) llm.ask("What changed in the EU AI Act this year?") # → cited, calibrated answer. # disputed & unknown facts flagged, never guessed.
# or just flip it in config model: llama3:8b vium: true
The moat
A stated 0.90 should be right about 90% of the time. Vium fits claim-level confidence against ground truth, marks what’s disputed or unknown, and refuses to be confidently wrong.
Refusing to be confidently wrong is the entire game.
Source reputation × recency decay.
Anti-slop: detect synthetic & syndicated copies.
Embed, cluster, count independent sources.
An LLM gate scored against held-out truth.
What we’re planning
The build is sequenced so the headline result — a 7B + Vium rivalling a 70B — is something we can measure early, not promise late.
The package schema, a single-source build, and the benchmark + calibration metrics from day one.
Authority → anti-slop → corroboration → calibrated judge. Known-false facts get dropped or flagged.
Package store, semantic cache, stale-refresh, and a background topic-builder.
Need-router and injection on Ollama, then the 7B-with-vs-without-Vium demo against the benchmark.
MCP server, SDK, cost-routing, and continuous evaluation.
Building the trust layer for an internet that machines now read more than people do.