A clearer view of AI.
THE INDEPENDENT AI.

HYPE OR FACT? · CHECKED 29 OCTOBER 2026

Osaurus + LFM3: a 1.7B model 'read the whole case file' offline and cited its sources

OUR VERDICT

Partly fact

Our original verdict: “Fact or hype: partly fact. LFM3 and offline use in Osaurus are real, but the demo used fictional sample files, searched them in chunks on a 64 GB Mac rather than reading the whole file, and its citations aren't shown being checked; LFM3's benchmarks are Liquid's own.”

THE CLAIM

“Journalists can't upload a leak to the cloud. So we turned off the Wi-Fi. A 1.7B model read the whole case file, followed the money, and cited its sources. LFM3. Day 0 on Osaurus. (Quoting Liquid AI's LFM3 launch, which says LFM3-1.7B is '#1 on the Pareto frontier of device intelligence'.)”

@OsaurusAI · Original post ↗

What’s actually true

  • The model is real and small. Liquid AI released LFM3 on 27 Oct 2026 as open weights. LFM3-1.7B is a text model with a 256K-token context that takes about 1 GB of memory at 4-bit. Ollama's download is 1.1-1.7 GB, and it's also on Hugging Face and in LM Studio.
  • 'Day 0 on Osaurus' is about right. Osaurus release 0.25.28 (28 Oct, about 13 hours after Liquid's launch post) added LFM3. Osaurus is a free, MIT-licensed Mac app that runs models on the Mac itself, so it does work with Wi-Fi off.
  • The video (37 seconds) shows Wi-Fi switched off, then an 'Investigative Journalist' agent running LFM3 1.7B (MLX 8-bit) on a folder called alder-bay-case. The model's own reasoning calls it 'fictional sample data', so this is a staged example, not a real leak.
  • 'Read the whole case file' is overstated. The agent made 29 steps with file-search and read-file tools, reading the folder in chunks, and took about 1 minute 31 seconds (about 1 minute 20 seconds is cut from the video). The folder's size is never shown.
  • 'Cited its sources' is partly shown. The answer gives file names and line numbers (for example council_minutes_2026-03-24.pdf, lines 17-19), but the video never opens those files, so we can't tell whether the citations are accurate.
  • The demo ran on a Mac with 64 GB of RAM. The model itself is small, and Liquid measured a full 128K-token context at 1.6 GB of extra memory, so an ordinary Apple Silicon Mac should run it. How well it copes with a large case file on an 8 GB Mac isn't shown.
  • Small models make mistakes, and Liquid says so. The model card says to 'use retrieval or search for knowledge-heavy tasks' and doesn't recommend it for complex multi-step reasoning. On Liquid's own tables it trails Qwen3.5-4B on 128K long-context recall (RULER 65.6 vs 84.6). For journalism, every citation still needs checking by a person.
  • The benchmarks are self-reported. Liquid ran the comparison tables on its blog and model cards, and the 'Pareto frontier of device intelligence' ranking is on pipette.liquid.ai, Liquid's own site. We found no independent test yet.
  • The privacy story is true offline, with caveats when online. In the Osaurus source code, crash reporting (Sentry) is on unless you switch it off, and the anonymous usage-data box (Aptabase) is pre-ticked at setup; the app also checks for updates. The privacy policy describes this data as opt-in and says chats, prompts, files and keys are not included. With Wi-Fi off, nothing can be sent.

Sources

Cost

Free to try. Osaurus is free and MIT licensed, with no account needed for local models (team Workspaces are $20 per workspace a month; cloud models are paid credits). LFM3 weights are free to download, but under the LFM Open License free commercial use is only for organisations under $10M annual revenue (counting affiliated companies); bigger companies need a licence from Liquid AI.

USE IT IF

Mac users with Apple Silicon (M1 or newer, macOS 15.5+) who want to ask an AI about private documents without uploading them, such as journalists, lawyers or small businesses, and are happy to check the answers themselves.

SKIP IT IF

Anyone who needs answers they can trust without checking, deep reasoning over a very large archive, or Intel Macs. For high-stakes work, use it to find where to look, then read the cited pages yourself.

Related AI tools

Related

How we knowChecked on 29 October 2026. Verdicts: Fact = the claim holds up; Partly fact = real, but with caveats or overstated parts; Hype = the headline claim isn't supported. Products change quickly, so check the maker's site before you buy. Our methods & sources ↗