Opens in a new tab
News

LFM2.5 on a workstation: count stored weights as well as active ones

In brief

Selecting fewer experts can save computation without shrinking the whole model to its active size. Follow loading, tool use and the complete local task.

5 min read

Is this the right starting point?
Before you read
Basic model-inference and agent-workflow concepts.
What you will take away
Distinguish active-parameter compute from full-model memory and evaluate a local model against the complete tool task.
Sources
Four blue, rust and gold chambers are stacked among rows of ivory alcoves, with colored ribbons connecting a circular hub on the left to a gold arrow on the right.
On this page4 sections

An engineer wants a workstation assistant to read an alert, find the runbook and suggest the next check. The model promises less calculation per piece of its answer. First, though, the workstation must load the whole file. Faster arithmetic helps only after the model fits somewhere the machine can reach it.

Liquid AI’s October 13 LFM2.5-24B-A2B release makes that distinction important. LFM2.5-24B-A2B has 24 billion learned numerical settings, called parameters, in total, with roughly 2.45 billion active for each token, a piece of input or output. Fewer active parameters can reduce computation; the others still need storage. Follow the hypothetical assistant from loading that model to completing the investigation.

A token selects four experts; the other weights still exist

The model card specifies 64 experts in each mixture-of-experts block and top-four routing. For a token, the router selects four sets of learned calculations rather than running all 64. A later token may select different ones.

This reduces per-token computation compared with a dense 24B model. It does not turn the checkpoint into a 2.45B file. Placement in fast memory, system RAM or a slower offload tier affects how quickly the selected weights become available. Storing those numbers with reduced numerical precision can shrink their memory footprint. That is quantization; its settings affect how much room the workstation needs.

The 131,072-token context brings other costs. Prefill processes input before the reply begins. As attention relates new tokens to earlier ones, it can reuse saved intermediate results called keys and values. Holding those results in a key-value cache avoids calculating them again at each step. More input adds work and memory. This reasoning-only model can also emit reasoning tokens before its final answer. Sparse routing does not remove those stages.

LFM2.5 has 24 billion total parameters and roughly 2.45 billion active per token, with a router selecting four experts from the larger stored pool.
Only selected parts compute each token, but the remaining weights need storage so a later token can select them.

Use the benchmark table to choose what to try

Liquid reports 184 decoded tokens per second on Apple M5 Max and 114 on AMD Ryzen AI Max+ 395 under its tested quantization, runtime and prompt conditions. These are vendor measurements, not universal speed guarantees.

Choosing a runbook lookup with the correct service name is one job. Repairing the diagnostic’s Python code is another. A tool-call score cannot settle both. The Berkeley Function Calling Leaderboard, or BFCL, evaluates tool calls; Liquid reports LFM2.5 and Qwen3.6-35B-A3B essentially level in its BFCLv4 comparison.

The rest of the table describes a mixed picture. LFM2.5 leads the compared models on AutomationBench and that Qwen model on ToolSandbox and LiveClaw. Qwen leads WebArena Verified browser work, GPQA Diamond knowledge questions, and SWE-Bench Verified and SciCode coding tests. Those are different published evaluations, not steps this workstation has already passed.

Liquid recommends retrieval or web search for knowledge-heavy work. For this engineer, inspect what follows the lookup: did the correct runbook arrive, and did the proposed diagnostic follow it? The call matters because of the work it enables.

A benchmark result belongs to the model, harness, settings and conditions producing it. Use the table to shortlist candidates, then compare representative work with relevant tools, hardware and acceptance checks held steady. The benchmark-and-harness guide explains when comparing whole setups is the useful choice.

Follow the workstation from cold load to useful answer

In a hypothetical triage task, the model reads a sanitized alert, retrieves a runbook, calls a read-only diagnostic and proposes the next check. A fast warm reply measures only part of that wait.

TrialHold constantObserveQuestion answered
Cold loadCheckpoint format and host stateLoad time, peak memory, page-ins, failuresWhether the full model fits the planned memory path
Warm tool loopHarness, tools, prompt and acceptance rubricTime to first token, dispatch latency, retries, accepted tasksWhether sparse execution improves the actual workflow
Long-context replayDocument set and arrival patternPrefill time, cache growth, throughput, latency percentilesWhether the 131K context is usable at the target objective
Failure recoveryInjected timeout and malformed tool resultRepeated actions, bounded failures, destination stateWhether the agent remains safe when a tool outcome is uncertain

Acceptance requires the appropriate diagnostic and evidence supporting the next step. Fluent prose cannot repair a wrong tool, repeated action or missing observation. The inference compatibility suite separately addresses protocol behavior.

Check where a local task still sends data

Local weights eliminate one cloud-model call, not every external connection. Retrieval, search, telemetry, package downloads and tools may still send data out. Follow the whole task before describing the agent as private.

Licensing also remains a separate question. Liquid uses LFM Open License v1.0. Its license page describes a $10 million annual-revenue commercial-use boundary, with larger entities needing a separate license. Open weights do not mean unrestricted enterprise permission.

The engineer needs a model file the serving program can read. Downloading it is not the same as running the assistant. With a supported GGUF build and compatible engine, for example, first establish that it loads. Then send the runbook lookup and inspect tool name and arguments. Loading checks the file pairing; the lookup checks application behavior.

The release lists native, FP8, GGUF, ONNX and MLX forms and documented routes through Transformers, vLLM, SGLang and llama.cpp. Use a documented pairing rather than treating the names as interchangeable. Record both versions and repeat the small lookup after upgrades. Model integrity, resource limits and serving-endpoint availability become maintenance work for the workstation’s operator.

Return to the engineer’s task. Can the workstation load the full model, read the supplied evidence and finish the diagnostic loop without unacceptable waiting? The smaller active path makes the trial worth considering. The larger stored model explains why its memory path still belongs in the result.

Source context

Source links appear within this article. Read them alongside the author’s analysis and evaluate the guidance against your environment.

Report an error or outdated detail

Keep reading

Where to go next