Quick wins for a faster PC:
Scan for outdated or missing drivers - takes under a minuteDriver Scan →Clear out junk files and repair common Windows errorsFree Scan →Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Some links on this page are affiliate links: if you buy through them we may earn a commission, at no extra cost to you.
LFM2.5-24B-A2B is a sparse Mixture-of-Experts (MoE) language model: it contains 24.0 billion parameters in total, but Liquid AI’s model card says only 2.45 billion are active for a token. That sparsity can reduce the computation needed to generate each token, which is why the model is positioned for on-device agents. It does not reduce the complete weight set that a runtime must store or load. In practical terms, LFM2.5 can compute more like a small model while carrying the storage and memory considerations of a much larger one. The sections below explain its routing design, memory implications, supported runtimes, published speed figures, capabilities, and the limits of the available evidence.
What does LFM2.5-24B-A2B’s MoE design actually mean?
Mixture-of-Experts divides parts of a neural network into multiple specialist “experts.” A router examines each token and sends it to only a few experts instead of evaluating every expert. LFM2.5-24B-A2B has 64 experts in each MoE block and uses top-4 routing, so four experts are selected for a token in those blocks. The model therefore has a large pool of learned weights, while each token activates a comparatively small slice of that pool. Liquid AI describes the architecture in its official model card.
As an Amazon Associate I earn from qualifying purchases.
The “24B” in the name refers to the total parameter count: 24.0 billion. The “A2B” shorthand points to approximately 2 billion active parameters, while the model card gives the more precise active figure of 2.45 billion. Active parameters are the weights involved in a token’s computation; they are not the amount of data that disappears from the checkpoint. This distinction is the key to understanding both the model’s speed claims and its local-memory requirements.
Why can a 24B model behave like a smaller model during generation?
For each generated token, the router avoids running all 24 billion parameters. Fewer matrix operations generally mean lower per-token compute than a dense 24B model, where nearly the entire parameter set participates in every token. That can improve decode speed and energy efficiency on hardware with suitable kernels and memory bandwidth.
Popular on Amazon
- 【90 Degree Full Function USB C Adapter】Supports the latest USB4 standard for up to 40Gbps of bandwidth, 8K video support, and up to 240W Fast charging; Support USB PD 3.1 Extended Power Range (EPR) including 140W, 180W and 240W. Your productivity is easily doubled. *Warm Tip: For the best fit, please make sure the USB-C port is fully exposed; thick or bulky cases may affect compatibility.
- 【Reliable Quality】Made from high-quality, scratch-resistant aluminum alloy, delicate metal frame. get excellent wear resistance and heat dissipation. Non-Slip Design for easier insertion and removal.
- 【Small and portable】Ultra thin size, ultra compact design, a petite physique that matches the minimalist design of your modern laptops. Coming with a slim figure that won't block the port next to the one it's plugged into.
- 【Compatibility】Widely work with most usb c, thunderbolt 4/3 devices. Compatibilit for ROG Ally, Steam Deck, Switch, MacBook Pro, Quest 3 VR headset, Surface Book 2, iPhone 18/17/16/15, Samsung Galaxy S23/S22/S21, Note 20/10, Google Pixel, ChromeBook, etc.
- 【Excellence Customer Support】If you experience any issues or cable failure, simply reach out to our dedicated customer service team. We provide a prompt response within 18 hours, providing you with the necessary technical support to resolve any concerns.
Sparsity is not a guarantee of a particular speed. Routing overhead, the implementation, quantization, prompt length, context management, and device memory bandwidth all influence results. A runtime still has to access the selected experts, and the complete checkpoint may need to be resident or efficiently paged. “2.45B active” should therefore be read as a compute characteristic, not as a universal statement that the model runs with the memory footprint of a 2.45B dense model.
What are the main architecture specifications?
Liquid AI’s model card lists a hybrid 40-layer design. Thirty layers use double-gated short-convolution blocks, while 10 use grouped-query attention (GQA). The MoE blocks contain 64 experts with top-4 routing. This combination is inherited from the earlier LFM2-24B-A2B family design, with LFM2.5 extending the context and vocabulary.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
| Specification | LFM2.5-24B-A2B | Local significance |
|---|---|---|
| Total parameters | 24.0 billion | Sets the scale of the full learned weight collection and checkpoint. |
| Active parameters | 2.45 billion (model-card figure) | Approximates the parameter workload selected per token; it is not a complete memory requirement. |
| Layers | 40 | 30 double-gated short-convolution layers and 10 GQA layers. |
| Experts per MoE block | 64 | Provides the specialist pool from which the router selects. |
| Routing | Top-4 | Four experts are selected for a token in an MoE block. |
| Maximum context | 131,072 tokens | Supports very long prompts, but increases cache and runtime memory demands. |
| Vocabulary | 128,000 tokens | Expanded from 65,000 in the earlier LFM2 design. |
The training description is also larger than the earlier model’s. Liquid AI says the pretraining corpus grew from about 17 trillion to about 33 trillion tokens; its model card gives a rounded 36-trillion-token training budget and the release mentions additional mid-training phases. These are Liquid AI’s reported figures, not independently audited totals, so they should be treated as publisher disclosures.
If only 2.45B parameters are active, why does the model still need substantial memory?
Because the inactive experts are still part of the model. A checkpoint must contain the weights for all experts, routing components, embeddings, attention and convolution layers, and other parameters. Depending on the runtime, those weights may all be loaded into system or video memory, or moved in a managed way. Quantization can reduce the bytes per weight, but it does not turn a 24B checkpoint into a 2.45B checkpoint.
Actual memory use depends on at least four variables:
Also popular on Amazon
- 5-in-1 Connectivity: Equipped with a 4K HDMI port, a 5 Gbps USB-C data port, two 5 Gbps USB-A ports, and a USB C 100W PD-IN port. Note: The USB C 100W PD-IN port supports only charging and does not support data transfer devices such as headphones or speakers.
- Powerful Pass-Through Charging: Supports up to 85W pass-through charging so you can power up your laptop while you use the hub. Note: Pass-through charging requires a charger (not included). Note: To achieve full power for iPad, we recommend using a 45W wall charger.
- Transfer Files in Seconds: Move files to and from your laptop at speeds of up to 5 Gbps via the USB-C and USB-A data ports. Note: The USB C 5Gbps Data port does not support video output.
- HD Display: Connect to the HDMI port to stream or mirror content to an external monitor in resolutions of up to 4K@30Hz. Note: The USB-C ports do not support video output.
- What You Get: Anker 332 USB-C Hub (5-in-1), welcome guide, our worry-free 18-month warranty, and friendly customer service.
- Format and precision: BF16, FP16, 8-bit, 4-bit and other formats use different amounts of storage and may have different kernel support.
- Checkpoint implementation: GGUF, MLX, ONNX and native Transformers-style checkpoints can have different overheads and loading behavior.
- Context length: the key/value cache and other runtime state grow as the prompt and generated conversation grow. A 131,072-token ceiling is a capability limit, not a promise that every device can use it comfortably.
- Runtime settings: batch size, parallel requests, offloading and allocator behavior can materially change peak memory.
The cited sources do not establish a universal minimum RAM or VRAM configuration. A device recommendation must therefore name the exact quantized checkpoint, runtime, context target and concurrency rather than infer a requirement from the 2.45B active count alone.
What changed from LFM2-24B-A2B to LFM2.5?
LFM2.5 retains the earlier hybrid architecture but extends the maximum context from 32,768 to 131,072 tokens—four times the earlier limit. The vocabulary grows from 65,000 to 128,000 tokens. Liquid AI also reports a substantially expanded pretraining scale and additional mid-training phases. Those changes affect what the model can ingest and how it was developed; they do not alter the basic MoE lesson that total and active parameters are different quantities.
Which local runtimes and formats are available?
Liquid AI lists several deployment paths: GGUF for llama.cpp, MLX for Apple Silicon, ONNX for cross-platform deployment, and native checkpoints for Transformers, vLLM and SGLang. The Hugging Face announcement presents these options as ways to run private agents on local or edge hardware.
Choose the path that matches your hardware and software stack:
- Apple Silicon: the MLX build is the specifically identified route for Apple hardware. Unified memory can be shared by CPU and GPU, but the available capacity still has to accommodate the full selected checkpoint, runtime overhead and context state.
- llama.cpp-compatible systems: GGUF is useful when you want a quantized file and a broad ecosystem of CPU/GPU back ends. Confirm that the build supports the model’s architecture and routing behavior.
- Cross-platform applications: ONNX can fit teams standardizing deployment across operating systems or accelerators, subject to an execution provider that supports the required operators.
- Server frameworks: Transformers, vLLM and SGLang expose the native checkpoint for development, batching and serving. Their memory and throughput behavior should be evaluated with the exact model revision and settings.
Before downloading, verify the model-card instructions, revision, tokenizer files, quantization compatibility and context setting. A format that loads successfully may still be impractical at the context length or concurrency you need.
The Tool Desk
Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →Outbyte Driver Updater FREEScan for outdated or missing drivers - takes under a minuteDriver Scan →Also popular on Amazon
- Sleek 7-in-1 USB-C Hub: Features an HDMI port, two USB-A 3.0 ports, and a USB-C data port, each providing 5Gbps transfer speeds. It also includes a USB-C PD input port for charging up to 100W and dual SD and TF card slots, all in a compact design.
- Flawless 4K@60Hz Video with HDMI: Delivers exceptional clarity and smoothness with its 4K@60Hz HDMI port, making it ideal for high-definition presentations and entertainment. (Note: Only the HDMI port supports video projection; the USB-C port is for data transfer only.)
- Double Up on Efficiency: The two USB-A 3.0 ports and a USB-C port support a fast 5Gbps data rate, significantly boosting your transfer speeds and improving productivity.
- Fast and Reliable 85W Charging: Offers high-capacity, speedy charging for laptops up to 85W, so you spend less time tethered to an outlet and more time being productive.
- What You Get: Anker USB-C Hub (7-in-1), welcome guide, 18-month warranty, and our friendly customer service.
How fast is LFM2.5-24B-A2B on edge hardware?
Liquid AI reports 184 tokens per second on an Apple M5 Max and 114 tokens per second on an AMD Ryzen AI Max+ 395, describing both as edge-inference measurements. These figures are company-reported results, not independent tests, and should be treated as reference points rather than guarantees for every laptop or memory configuration.
The release says the comparison used Liquid AI’s Pipette, SGLang 0.5.22, a 1,024-token input, up to 256 output tokens and BF16. Changing any of those conditions—including quantization, prompt length, thermal limits, power mode or software version—can change throughput. Token-per-second numbers are also most meaningful when comparing the same prompt, output length, precision, batch and concurrency.
Liquid AI separately reports more than 21,000 output tokens per second on one NVIDIA B300 at high concurrency. That is a server-serving result, not a laptop or ordinary single-user on-device result. It illustrates how batching and high-end accelerators can multiply aggregate throughput, but it should not be used to predict interactive performance on a consumer computer.
What kinds of tasks is the model intended to handle?
Liquid AI recommends LFM2.5-24B-A2B for agents that need to reason, call tools and run reliably on the device. The company reports that the model leads or matches the compared systems on five of eight tool-use and agentic benchmarks, and reports leadership on the three instruction-following benchmarks in its table.
Do these 3 things before closing this tab:
1Clear out junk files and repair common Windows errors2Fix the driver behind crashes, sound loss and screen glitches3Repair Windows errors before they cause bigger problemsOne reported example is AutomationBench, where LFM2.5 scores 27.69 versus 8.72 for the next listed model, Qwen3.6-35B-A3B. The comparison is from Liquid AI’s evaluation, run with vLLM and benchmark-recommended generation settings. It is evidence about that published setup, not a universal ranking for every agent framework or prompt.
Also popular on Amazon
- Portable and powerful USB-C HUB: BENFEI USB Type-C HUB, with super-soft and knot-free silicone woven design cable, meets most mobile office needs. Compact, lightweight, stylish, and powerful portable USB C Hub equipped with 1 x HDMI port, 1 x 100W charging, and 3 x USB ports. 18-month warranty, 24-hour response, to ensure you feel at ease when using our product.
- Design centered on comfort and reliability: Thanks to BENFEI's end-to-end in-house cable production capability, in-house PCBA and assembly capability, using the industry's most advanced silicone woven design and process, 20cm cable in length, no knots, super-soft, the HUB is easy to use in all scenarios: laptop, tablet, stand etc. Super-soft, 25000+ life cycles, to meet your daily carrying and office needs.
- 100W Charging: Support up to 90W USB C pass-through charging via Type-C port to keep your laptop powered. 10W is reserved for other interface operations. No data and video function on the Type-C port.
- 4K HDMI Display: The HDMI port supports media display at resolutions up to 4K 30Hz, keeping every incredible moment detailed and ultra vivid. Please note that the C port of the Host device needs to support video output.
- Transfer Files in Seconds: Transfer files and from your laptop at speeds up to 10 Gbps with USB A 3.2 port. Extra 2 USB A 2.0 ports are perfectly for your keyboards and mouse.
The same evaluation shows larger competitors doing better on several knowledge-heavy and coding measures, including GPQA Diamond, HLE, SWE-Bench Verified (OpenCode) and SciCode. Liquid AI’s model card therefore gives a practical warning: the model is not recommended for knowledge-heavy tasks on its own. Pairing it with retrieval or web search can supply current facts while allowing the model to focus on tool calling and orchestration.
How should you decide whether it fits your device?
Start with the workload, then size the deployment. A sensible decision process is:
- Define the interaction: estimate prompt length, response length, number of simultaneous users and whether the model must call tools repeatedly.
- Select a format: choose a supported GGUF, MLX, ONNX or native checkpoint and record its precision or quantization level.
- Budget memory for the whole system: include all model weights, embeddings, runtime overhead, operating-system use, context cache and any other models or tools sharing the device.
- Set a realistic context target: do not size hardware solely for the 131,072-token maximum if normal sessions are much shorter; conversely, long documents can make cache memory the limiting factor.
- Check software support: confirm the model revision, tokenizer, accelerator backend and routing kernels work together in the chosen runtime.
- Compare like with like: use the same input/output lengths, precision, batch size and concurrency when interpreting published or internal throughput numbers.
- Evaluate your actual task: test tool calls, structured output, retrieval quality, coding or domain prompts with an evaluation set representative of your application. Benchmark labels alone cannot predict every workflow.
An Apple Silicon laptop is a plausible example because Liquid AI publishes an M5 Max result and provides an MLX path, but the sources do not identify a minimum RAM configuration or establish that result across all MacBook variants. AMD Ryzen AI Max+ systems are another named edge category, while a B300 belongs to high-concurrency server inference rather than ordinary personal use.
Outdated Drivers Are Slowing You Down
One free scan finds every outdated or missing driver and matches the right update for your exact hardware.Free scan · exact hardware matchPC Slower Than It Used to Be?
A free scan shows the junk files, broken settings and background clutter dragging Windows down - then fixes them in one click.Free scan · Windows 10 & 11What are the main trade-offs of sparse activation?
| Potential advantage | What it means | Important limitation |
|---|---|---|
| Lower per-token compute | Only selected experts process each token. | Routing and memory access still add overhead; speed depends on implementation. |
| Large total capacity | Different experts can specialize in different patterns. | The full 24B weight collection still affects storage and loading. |
| Long context | Up to 131,072 tokens are supported by the model card. | Long context consumes more cache memory and can reduce practical throughput. |
| Local privacy and availability | Inference can run without sending prompts to a remote API when the device is configured locally. | Local deployment shifts responsibility for updates, hardware capacity and security to the operator. |
| Agent and tool focus | Published results emphasize tool use and instruction following. | Knowledge-heavy and some coding benchmarks favor larger models; retrieval may be needed. |
How should LFM2.5 be compared with another local MoE model?
Compare total parameters and active parameters separately, then compare the actual memory footprint of the selected checkpoint. Next align context length, quantization, prompt and output sizes, decode speed, batch or concurrency, tool-use scores and performance on your own tasks. A model with fewer active parameters may still require more storage than a dense model with a similar active compute budget, while a model with a shorter context may be easier to run for document-heavy work.
Published benchmark tables are useful directional evidence, but they are not a universal leaderboard. Liquid AI’s figures come from its stated evaluation setup, and the Hugging Face article republishes release details rather than independently replicating them. The LFM2 technical report provides background on the earlier LFM2 edge-oriented hybrid lineage; it should not be used to substitute earlier specifications for the LFM2.5-24B-A2B model.
Also popular on Amazon
- 【Read before You Buy】 Note: This adapter is USB 2.0 version and cannot transmit HDM, VGA, or other video signals. Nevertheless, they are fully capable of charging devices,duplicating files, and can be seamlessly integrated with USB-C earphones, wireless network adapters, and SD card adapters.
- 【Easy to Use and Safe】Thanks to modern manufacturing techniques, we've made these adapters sufficiently compact so that they can remain inserted in your device without noticeable protrusion or inconvenience. This means you don't have to constantly plug and unplug them, reducing the likelihood of wear and misalignment, and making them more convenient to carry around.
- 【Reliable Quality】Engineered for enhanced durability, durable aluminum alloy shell material,get excellent wear resistance and heat dissipation,the male USB C and female USB connectors & the male USB and female USB C connectors are designed to be plugged and unplugged up to 20000 times. Provides data transfer, device charging, memory space expansion, and additional computer peripherals.
- 【Think more for you】Do you need to connect a USB-C-enabled device to an odd device that only has USB-A ports? Then these USB-A male to USB-C female adapters are made for you.Simply insert the adapter into the USB-A port on the device, and you can instantly connect your USB-C device to the device that originally didn't have the correct connection port.
- 【Trusted xiwxi Quality and Service】As a responsible brand seller, you will receive xiwxi's 12-month Quality assurance and lifetime friendly customer service, you're not just buying a product,you're investing in peace of mind with all-round quality and exceptional customer service.
What is the bottom-line implication for on-device AI?
LFM2.5-24B-A2B demonstrates the central bargain of MoE design: a model can offer a 24B-scale pool of parameters while activating only about 2.45B for each token. That can make interactive local inference more attainable than a dense 24B model, especially on optimized edge runtimes. The trade-off is that the complete model still has to be stored, loaded or managed, and long context can add substantial runtime memory.
For an on-device agent that calls tools, follows instructions and benefits from local privacy, the architecture and Liquid AI’s published edge results are relevant. For fact-heavy answers, retrieval or web search remains important. Treat the vendor’s speeds and benchmark scores as configuration-specific reference data, and choose hardware only after accounting for the exact format, context and concurrency your application requires.
Frequently Asked Questions
Does 2.45B active parameters mean I can run LFM2.5 in 2.45B of RAM?
No. Active parameters describe the weights used for a token’s computation. The runtime still needs the full 24B parameter collection, plus quantization metadata, working memory and context cache; the required capacity varies by format and settings.
Can LFM2.5-24B-A2B work without an internet connection?
Its supported local formats allow inference on a configured device, so the model itself need not send prompts to a hosted service. An offline agent will not automatically have current information, however; any retrieval source or tool it calls must also be available locally.
Is a 131,072-token context practical on every computer?
No. The model card states the maximum context length, but using very long prompts increases cache memory and can lower throughput. Practical context limits depend on precision, runtime, available memory and the amount of other work the device is doing.
Why might a larger dense model score better even if LFM2.5 uses tools efficiently?
Tool-use and knowledge recall are different capabilities. Liquid AI reports stronger results for some larger models on knowledge-heavy and coding benchmarks, while positioning LFM2.5 for agentic orchestration; retrieval can help close the factual gap.
Which file format should I choose for a laptop deployment?
Use the format supported best by your hardware and runtime: MLX is the identified path for Apple Silicon, GGUF targets llama.cpp ecosystems, ONNX supports cross-platform execution providers, and native checkpoints fit Transformers, vLLM or SGLang workflows. Verify revision and quantization compatibility before committing.

