Good morning — it's Thursday, October 29, 2026. The lead today is LFM3-1.7B, Liquid AI's tiny local-agent model. From the bench: the same model's scoreboard, against a rival five times its size. Plus the Qotom ten-port box's measured catch, an OCR model built for hundred-page documents, and Immich's people sharing. Don't keep us a secret: forward this email to a fellow home labber, or share the web version.
Share this issue: X WhatsApp Reddit Facebook LinkedIn Text Email
LFM3-1.7B, released by Liquid AI on October 23 and trending on Hugging Face this week, is a 1.73B text model built for one job: running local agents. Its tri-hybrid design mixes gated convolutions with sliding-window and global attention across 24 layers, carries a 256K-token context, and — at 4-bit precision — takes about 1 GB of memory (vendor). It was post-trained for agent harnesses, tool calling, and structured outputs.
On Liquid AI's own numbers (vendor), it out-scores Qwen3.5-9B — a model five times its size — on tool calling and agent work, and loses the coding test; the full scoreboard is on the bench below. The honest limits are Liquid AI's own: use retrieval for knowledge-heavy work, and don't reach for it for agentic coding. The license matters too: the LFM Open License v1.0 covers commercial use only below $10M annual revenue. For your rack: a tool-using sidecar that runs anywhere.
LFM3-1.7B is a tool-user, not a coder — Liquid AI says so itself, which is why today's head-to-head is not the usual coding leader. It is Liquid AI's own table, against Qwen3.5-9B, a model five times its size.
Verdict: on Liquid AI's numbers, the 1.7B wins the agent and tool work and loses the coding test. Higher is better on every line; LFM3-1.7B's score is first.
BFCL V4 (tool calling): 71.39 for LFM3-1.7B, 62.83 for Qwen3.5-9B — LFM3 ahead by 8.56.
LiveClaw (agents): 45.60 for LFM3-1.7B, 35.77 for Qwen3.5-9B — LFM3 ahead by 9.83.
IFEval (instruction following): 96.66 for LFM3-1.7B, 92.08 for Qwen3.5-9B — LFM3 ahead by 4.58.
Terminal-Bench 2.1 (coding): 12.92 for LFM3-1.7B, 17.98 for Qwen3.5-9B — Qwen ahead by 5.06.
Evidence: every score is Liquid AI's own run (vendor), all models served in vLLM at full context. No independent replication has appeared yet.
Setup notes: GGUF builds run in llama.cpp, Ollama, and LM Studio; an MLX build covers Apple Silicon. Liquid AI puts the 4-bit footprint at about 1 GB (vendor). Difficulty tag: drop-in.
Ten ports, one measured catch. ServeTheHome's review of Qotom's Q51551UT puts a number on the shared-bandwidth caveat from No. 002's drops (independent): its two copper 10GbE ports together cap at about 12Gbps, 11 to 12 in their tests, because the 16-year-old X540 sits on a ~16Gbps link and cannot step down to 2.5 or 5Gbps. The SFP+ pair is the full-speed path. Package power stays modest: 8.1W idle, about 28W steady under load.
An OCR model that reads the whole binder. Baidu released Unlimited-OCR-2 on October 27: a 3.34B, MIT-licensed model whose sliding-window attention holds a fixed set of working references — 512 tokens, on the card's example — so memory stays nearly flat as documents grow to hundreds of pages (vendor). Serving is the catch: SGLang needs a customized wheel and remote code. Fiddly today; exactly the shape a scan archive wants tomorrow.
Immich learns to share people. October's Immich v3.3.0 adds people sharing, birthday memories, auto-stacked mobile edits, and OAuth claims synced on every login. Updating? Go straight to v3.3.5, the current patch, which also carries the CA-compromise fix covered in No. 003. The quiet upgrade in 3.3.0: set MACHINE_LEARNING_MODEL_REVISION=v2 and the re-exported models run faster on less memory, with no jobs to re-run.
Thursday, 9:40 PM. Your rack UPS fails its self-test and reports a weak battery, with storms forecast for tomorrow. Everything is still running on mains power, your graceful-shutdown setup was tested last month, and no spare battery is on the shelf.
Your call:
A. Run a full discharge test tonight to learn the battery's real runtime
B. Leave the UPS in service, set shutdown to trigger earlier, and order a battery tonight
C. Shut the whole rack down until the replacement battery arrives
100 Hero Points for the right call. Your score is on the outcome page, ready to share.
It's easier to ask forgiveness than it is to get permission.
— Grace Hopper
Reading this because a fellow home labber forwarded it? Get your own copy every morning — subscribe free, and pass it on — that is how this grows.
How we source: facts come from primary sources opened for this issue. Benchmarks are labeled (vendor), (independent), or (community), never blended. Deal prices are checked the day they run. This is not buying advice.
Corrections: none today.
Reader-supported: paid subscriptions, labeled sponsors, and affiliate links on some gear and deal items (we may earn a commission). Coverage is never paid for — sponsors and affiliate programs don't pick stories.