Key fact: LFM2.5-8B-A1B at Q4 needs 7.0 GB. Q8 needs 12.0 GB. FP16 needs 21.9 GB.
How we got it: 8.3 × 0.5 × 1.2 + 2 = 6.98, rounded to 7.0. 8.3 × 1.2 + 2 = 12.0. 8.3 × 2.4 + 2 = 21.9. Total parameters, not the 1.5B active count.
Checked:
LFM2.5-8B-A1B Hardware Requirements: VRAM and GPU Guide
By LLM Hardware team · Methodology ·
VRAM figures are computed from the published 8.3B total parameter count using the sitewide methodology. Tokens/sec estimates scale with the 1.5B active parameter set and the listed memory bandwidth.
Key facts
- Model
- LiquidAI/LFM2.5-8B-A1B
- Shape
- 8.3B total parameters, 1.5B active
- 8B Q4
- 8.3 × 0.5 × 1.2 + 2 = 7.0 GB. An 8 GB GPU fits that estimate.
- 24B Q4
- LFM2.5-24B-A2B is 24.0B total. Q4 is 16.4 GB and misses 16 GB. Q8 is 30.8 GB.
- License
- lfm1.0 — read it before commercial use
Checked 2026-10-13. Author: LLM Hardware team. This site does not store or show Amazon prices. Check current price on Amazon.
Updated 24 October 2026 · LLM Hardware team · 8.3B and 24.0B MoE · d1-3B decision model · llama.cpp / MLX
Citeable facts
Key fact: LFM2.5-24B-A2B is 24.0B total and 2.45B active. Q4 is 16.4 GB and misses 16 GB. Q8 is 30.8 GB and fits 32 GB with about 1 GB left.
How we got it: 24 × 0.5 × 1.2 + 2 = 16.4. 24 × 1.2 + 2 = 30.8. 24 × 2.4 + 2 = 59.6 for BF16. HEAD Content-Length on LiquidAI/LFM2.5-24B-A2B-GGUF on 2026-10-13: Q4_K_M 14,526,258,464 bytes, Q5_K_M 17,029,602,592 bytes, Q8_0 25,493,732,640 bytes, BF16 47,962,094,880 bytes. Command: llama-cli -hf LiquidAI/LFM2.5-24B-A2B-GGUF. There is no ollama run tag. ollama run lfm2.5 is the older 8B library tag.
Checked:
Key fact: Liquid d1-3B at Q4 needs 3.9 GB. The official local command is llama-server, and this page has no ollama run tag for it.
How we got it: Card total 3.12B. 3.12 × 0.5 × 1.2 + 2 = 3.9. Q8 is 5.7 GB. BF16 formula is 9.5 GB. Hugging Face GGUF files on 12 October 2026, not re-fetched on 13 October: Q4_K_M 1,674,456,672 bytes, Q8_0 2,874,781,280 bytes, BF16 5,403,160,160 bytes. The Q8 mmproj is 583,109,728 bytes and is a separate file. Command from the GGUF card: llama-server -hf LiquidAI/d1-3B-GGUF:Q8_0.
Checked:
Methodology — parameter math, quantization bytes, and the source list.
Liquid AI shipped LFM2.5-8B-A1B on 28 May 2026 — a reasoning Mixture-of-Experts with 8.3B total parameters and 1.5B active per token. Q4 on this site is 8.3 × 0.5 × 1.2 + 2 = 7.0 GB, so an RTX 4060 8GB can load it with about 1 GB left. On 13 October 2026 Liquid published LFM2.5-24B-A2B, 24.0B total and 2.45B active. Its Q4 is 16.4 GB, so a 16 GB card does not run it. On 7 October 2026 Liquid also published d1-3B, a 3.12B decision model trained from LFM2.5-VL-3B. It answers typed questions in one forward pass and does not write chat text. Its Q4 estimate is 3.9 GB.
What is LFM2.5-8B-A1B?
LFM2.5-8B-A1B is Liquid AI's on-device reasoning model. It builds on the LFM2-8B-A1B release from late 2025 with a doubled vocabulary, scaled-up pretraining (12T to 38T tokens), an explicit chain-of-thought before every answer, and a 128K context window.
- Architecture Mixture-of-Experts. 8.3B total parameters across experts, 1.5B active per token. Speed of a 1.5B model, knowledge depth of an 8B.
- Context 128K tokens. Long enough for full codebases, long documents, or multi-turn agent traces.
- Reasoning Always emits an explicit chain-of-thought before its final answer. Treat the <think> block as internal scratchpad and present only the post-think output to end users.
- Hugging Face LiquidAI/LFM2.5-8B-A1B (base + instruct + GGUF + MLX). License: lfm1.0.
LFM2.5-8B-A1B VRAM by Quantization
RTX 4070 12 GB on AmazonMoE models still load every expert into VRAM at inference time. The figures below use the sitewide formula: VRAM = params × bytes_per_param × 1.2 + 2. Q4_K_M is 0.5 bytes, Q8 is 1.0, FP16 is 2.0. The 1.2 factor is a short context, not the 128K window.
| Quantization | VRAM (8.3B total) | Fits on | Notes |
|---|---|---|---|
| Q4_K_M | 7.0 GB | RTX 4060 8GB, about 1 GB left | 8.3 × 0.5 × 1.2 + 2. Short context only on 8 GB. |
| Q5_K_M | 8.2 GB | 12 GB GPU | 8.3 × 0.625 × 1.2 + 2. Misses an 8 GB card. |
| Q8_0 | 12.0 GB | 16 GB comfortable; 12 GB is the edge | 8.3 × 1.2 + 2. A 12 GB card has no extra KV room. |
| FP16 | 21.9 GB | 24 GB GPU | 8.3 × 2.4 + 2. A 16 GB card misses it. |
Want a context-aware estimate? The VRAM Calculator projects KV-cache growth for the 128K window.
Why The Active vs Total Distinction Matters
LFM2.5-8B-A1B is one of a wave of on-device MoE releases, and its memory and speed behaviour is unlike a dense 8B model.
VRAM is set by TOTAL params
All 8.3B parameters must be resident in VRAM because routing happens per-token and any expert can be selected. Plan VRAM as if it were a dense 8B.
Speed is set by ACTIVE params
Only 1.5B parameters fire per forward pass. Tokens/sec on the same hardware are 3-5x what you would see from a dense 8B model.
Great for low-power devices
Liquid AI quotes ~30 tok/s on a phone and 253 tok/s on M5 Max. That speed comes from the small active set, not the total size.
Long context is cheap
KV-cache scales with active params for the routed FFN block, so the 128K window is more affordable in memory than on a comparable dense 8B.
What LFM2.5-8B-A1B Can You Run on Your GPU?
Find your GPU below. Each card shows which quantization tier fits and what to watch out for.
RTX 4060 8GB
Runs:
- +Q4_K_M at 7.0 GB, about 1 GB left
- +Short context only
Does not fit:
- -Q5_K_M at 8.2 GB
- -Q8_0 at 12.0 GB
- -FP16 at 21.9 GB
8.3 × 0.5 × 1.2 + 2 = 7.0. Liquid’s 1.5B active count is why this card can feel quick. This page does not publish a new tokens-per-second measurement.
Intel Arc B580 12GB
Runs:
- +Q4_K_M at 7.0 GB
- +Q5_K_M at 8.2 GB
Does not fit:
- -Q8_0 at 12.0 GB is the edge, with no KV room
- -FP16 at 21.9 GB
Stock was not confirmed. Confirm a llama.cpp Vulkan or SYCL build before you buy. Check current price on Amazon.
RTX 4070 12GB
Runs:
- +Q4_K_M at 7.0 GB
- +Q5_K_M at 8.2 GB
Does not fit:
- -Q8_0 at 12.0 GB is the edge
- -FP16 at 21.9 GB
Bandwidth is 504 GB/s. Check current price on Amazon.
RTX 4060 Ti 16GB
Runs:
- +Q4_K_M, Q5_K_M, and Q8_0 (12.0 GB)
Does not fit:
- -FP16 at 21.9 GB
This row is the 8B model. LFM2.5-24B-A2B Q4 is 16.4 GB and does not fit 16 GB. Check current price on Amazon.
RTX 4090 24GB
Runs:
- +Q4 through Q8
- +FP16 at 21.9 GB
Does not fit:
- -Nothing in the LFM2.5-8B-A1B weight set misses 24 GB at these quants
24 GB fits LFM2.5-24B-A2B Q4 (16.4 GB) and Q5 (20.0 GB). Q8 is 30.8 GB and does not fit. Listed bandwidth is 1,008 GB/s. ASIN B0BG94PS2F was unavailable on 13 October 2026.
AMD RX 7900 XTX 24GB
Runs:
- +Q4 through Q8
- +FP16 at 21.9 GB
Does not fit:
- -Nothing in the LFM2.5-8B-A1B weight set misses 24 GB at these quants
24 GB at 960 GB/s. Confirm a ROCm or Vulkan llama.cpp build. Check current price on Amazon.
Mac mini M4 16GB
Runs:
- +Q4_K_M at 7.0 GB after a macOS reserve
- +Q5_K_M at 8.2 GB is tight
Does not fit:
- -Q8_0 at 12.0 GB after a ~4 GB reserve
- -FP16 at 21.9 GB
Usable memory is about 12 GB after the reserve. Liquid quotes about 253 tok/s on an M5 Max. That figure is for an M5 Max, not an M4.
Mac mini M4 24GB
Runs:
- +Q4, Q5, and Q8 (12.0 GB) after a ~4 GB reserve
Does not fit:
- -FP16 at 21.9 GB. About 20 GB is usable, so 21.9 GB misses
The 8B Q8 fits. LFM2.5-24B-A2B Q4 is 16.4 GB and is tight after a ~4 GB reserve. Q8 is 30.8 GB and does not fit. Check current price on Amazon.
Inference Speed by Hardware
The tokens-per-second column is a bandwidth sketch from the formula under the table. Liquid’s published d1 latency (8 ms on an RTX 4090, 30 ms on an M5 Pro) is a decision-model figure, not a row in this table.
| Hardware | Bandwidth | Q4 tok/s | Q8 tok/s | Notes |
|---|---|---|---|---|
| RTX 4090 24GB | 1,008 GB/s | ~180 t/s | ~120 t/s | Bandwidth ceiling, not compute |
| RTX 4070 12GB | 504 GB/s | ~95 t/s | ~62 t/s | Strong sweet-spot pick |
| RTX 4060 Ti 16GB | 288 GB/s | ~55 t/s | ~36 t/s | Best 16 GB value |
| Intel Arc B580 12GB | 456 GB/s | ~85 t/s | ~55 t/s | Budget winner |
| RTX 4060 8GB | 272 GB/s | ~50 t/s | — | Q4 only |
| Mac mini M4 Pro 24GB | 273 GB/s | ~50 t/s | ~33 t/s | MLX backend |
| Mac mini M4 16GB | 120 GB/s | ~22 t/s | ~14 t/s | Q4 recommended |
Estimate: tokens/sec ≈ bandwidth (GB/s) / (active_params × bytes). The 1.5B active set keeps numerator low — speeds run well above what dense 8B models hit on the same cards.
How to Run LFM2.5-8B-A1B Locally
llama.cpp / Ollama
ollama run hf.co/LiquidAI/LFM2.5-8B-A1B-GGUF This command is for LFM2.5-8B-A1B, not d1. The Q4 formula is 7.0 GB. d1-3B uses llama-server -hf LiquidAI/d1-3B-GGUF:Q8_0 and has no ollama run tag on this page.
MLX (Apple Silicon)
mlx_lm.generate --model LiquidAI/LFM2.5-8B-A1B Day-one MLX support. Best path on Mac for both performance and the 128K context window. Use the MLX 4-bit quant for the M-series mini class machines.
vLLM / SGLang
vllm serve LiquidAI/LFM2.5-8B-A1B For server deployments or high-concurrency on-prem use. Liquid reports 18,500 tokens/sec at high concurrency on a single H100.
For installation walkthroughs, see how to run LLMs locally and the Ollama cheat sheet.
Which Hardware Should You Buy for LFM2.5-8B-A1B?
Budget — 8 GB
RTX 4060 8GB
Q4 is 7.0 GB, so about 1 GB remains on 8 GB. Q8 at 12.0 GB does not fit. d1-3B Q4 at 3.9 GB does fit. Check current price on Amazon.
Best value 12 GB
Intel Arc B580 12GB
Q8 is 12.0 GB, which is the edge of this card, not spare room. Q4 at 7.0 GB fits with room. Confirm llama.cpp Vulkan/SYCL support. Check current price on Amazon.
CUDA mid-range
RTX 4070 12GB
Q4 and Q5 fit. Q8 at 12.0 GB fills the card. Check current price on Amazon.
16 GB sweet spot
RTX 4060 Ti 16GB
Comfortable Q8 with room for the full 128K context KV cache. The best mid-range NVIDIA pick for long-context use.
On-device Mac
Mac mini M4 24GB
MLX support landed day one. 24 GB unified memory holds Q8 plus the 128K KV cache. Silent and fast — Liquid AI quotes ~253 tok/s on M5 Max.
24 GB FP16
AMD RX 7900 XTX 24GB
24 GB at 960 GB/s for FP16-quality inference. Cheaper than an RTX 4090 with the same VRAM. ROCm and Vulkan both work with current llama.cpp builds.
For the full cross-budget GPU comparison, see the best GPU for LLMs guide.
LFM2.5-24B-A2B does not fit 16 GB
On 13 October 2026 Liquid published LFM2.5-24B-A2B under the lfm1.0 license. The card lists 24.0B total parameters, 2.45B active, 40 layers, 64 experts with top-4 routing, a 131,072-token context, and a 128k vocabulary. It is a reasoning model that always thinks. It is not a chat-only knowledge model; pair it with retrieval. Count the 24.0B total, not the 2.45B active count. Q4 is 24 × 0.5 × 1.2 + 2 = 16.4 GB. Q5 is 24 × 0.625 × 1.2 + 2 = 20.0 GB. Q6 is 24 × 0.75 × 1.2 + 2 = 23.6 GB. Q8 is 24 × 1.2 + 2 = 30.8 GB. BF16 is 24 × 2.4 + 2 = 59.6 GB. A 16 GB card misses Q4. A 24 GB card fits Q4 and Q5 and misses Q8. A 32 GB card fits Q8 with about 1.2 GB left and misses BF16. A 128 GB unified box fits the BF16 formula.
| Quant | Formula | File bytes on 13 Oct | Fit |
|---|---|---|---|
| Q4_K_M | 16.4 GB | 14,526,258,464 | Misses 16 GB. Fits 24 GB. Tight on a 24 GB Mac after a ~4 GB reserve. |
| Q5_K_M | 20.0 GB | 17,029,602,592 | Fits 24 GB. Misses a 24 GB Mac after the reserve. |
| Q8_0 | 30.8 GB | 25,493,732,640 | Misses 24 GB. Fits 32 GB with about 1 GB left. Edge of a 36 GB Mac after the reserve. |
| BF16 | 59.6 GB | 47,962,094,880 | Misses 32 GB. Fits 128 GB unified, including the EVO-X2. |
File bytes are Content-Length on LiquidAI/LFM2.5-24B-A2B-GGUF on 13 October 2026. The Q4_K_M file is smaller than 16.4 GB. Our VRAM formula still adds the 1.2 factor and 2 GB, so a 16 GB card is not a fit. The GGUF README command is llama-cli -hf LiquidAI/LFM2.5-24B-A2B-GGUF --temp 0.1 --top-k 50 --repeat-penalty 1.05. The QAD file is llama-cli -hf LiquidAI/LFM2.5-24B-A2B-GGUF --hf-file LFM2.5-24B-A2B-QAD-Q4_0.gguf --temp 0.1 --top-k 50 --repeat-penalty 1.05. That QAD-Q4_0 file was 13,578,190,880 bytes. MLX weights are LiquidAI/LFM2.5-24B-A2B-MLX-4bit. LM Studio is listed on the card. Ollama stable v0.40.3 and the v0.40.4-rc1 prerelease do not add an ollama run tag for this 24B model. ollama run lfm2.5 is the older 8B library tag.
Liquid’s own Pipette note, llama.cpp Q4_K_M at 4K context, quotes 184 tok/s on an Apple M5 Max and 114 tok/s on an AMD Ryzen AI Max+ 395. The card also quotes up to 21K output tok/s on one NVIDIA B300 with SGLang BF16. Those figures are Liquid’s. The 8B speed table above is a bandwidth sketch for the 8B model only. Do not read those rows as 24B speeds. Size the 24B on the VRAM calculator, then compare the Apple Silicon buy list and the Strix Halo mini PCs.
Liquid d1-3B is not a chat model
On 7 October 2026 Liquid published d1-3B and d1-omni-600M. d1-3B is trained from LFM2.5-VL-3B. The card lists 3.12B total parameters, a 400M SigLIP2 vision encoder inside that total, a 32,768-token context, and Apache-style use under the LFM license file. It returns a yes/no, a named choice, or a score in one forward pass. Output tokens are zero. It does not replace ollama run for writing or coding.
| Model | Card total | Q4 formula | What to run |
|---|---|---|---|
| d1-3B | 3.12B | 3.9 GB | llama-server -hf LiquidAI/d1-3B-GGUF:Q8_0, then POST /v1/systemone. Q8 formula is 5.7 GB. Fits 8 GB. Q4_K_M file is 1,674,456,672 bytes. |
| d1-omni-600M | 587M | 2.4 GB | transformers>=5.15, trust_remote_code. Text plus image, or text plus up to 30 s of audio, not both. Liquid does not publish a speed number. Fits 8 GB. |
Liquid’s own d1-3B card lists one-question latency of 8 ms on an RTX 4090 and 30 ms on an Apple M5 Pro, in bf16, median of 20 warm runs. Those are decision latencies, not generative tokens per second. Mistral Large 4 weights were still not downloadable on 12 October 2026, and this pass did not recheck that repo. Mistral’s announcement says the weights drop at the end of the month, and the Hugging Face upcoming page lists 31 October 2026. This guide does not size a GPU for it. The sibling pages are the VRAM calculator, the Ollama cheat sheet, and the Mistral hardware guide.
Frequently Asked Questions
How much VRAM does LFM2.5-8B-A1B need?
LFM2.5-8B-A1B has 8.3B total parameters. Q4 is 8.3 × 0.5 × 1.2 + 2 = 7.0 GB. Q8 is 12.0 GB. FP16 is 21.9 GB. An 8 GB GPU fits the Q4 estimate with about 1 GB left. A 12 GB GPU is the edge of the Q8 estimate. LFM2.5-24B-A2B is a separate 24.0B model. Its Q4 is 16.4 GB and misses 16 GB. Its Q8 is 30.8 GB. Liquid d1-3B is a separate 3.12B decision model: Q4 is 3.9 GB and it is not a chat model.
Can an 8GB GPU run LFM2.5-8B-A1B?
Yes, at Q4. 8.3 × 0.5 × 1.2 + 2 = 7.0 GB, so an RTX 4060 8GB has about 1 GB left for a short context. Q8 is 12.0 GB and misses 8 GB. The model activates 1.5B parameters per token. This page does not add a new tokens-per-second measurement for that speed.
What is the difference between total and active parameters?
Total parameters (8.3B) is the full size of the model and determines VRAM. Active parameters (1.5B) is the number used per forward pass and determines inference speed. LFM2.5-8B-A1B routes each token through a small subset of experts, so it has the speed of a 1.5B model with the knowledge depth of an 8B model — the core trade-off of MoE architectures.
Does LFM2.5-8B-A1B run on Apple Silicon?
Yes. Liquid AI shipped day-one MLX support, and Liquid reports about 253 tokens/sec on a Mac with M5 Max. A Mac mini M4 16GB fits the model at Q4_K_M with plenty of headroom for the 128K context window. For Q8 quality, the Mac mini M4 24GB is the comfortable pick. Apple unified memory works well for MoE models because the entire expert set must stay resident regardless of which GPU you use.
Is LFM2.5-8B-A1B open source?
The model weights are openly downloadable from Hugging Face under the lfm1.0 license. It is best described as open-weight rather than fully open-source — review the Liquid AI license terms before commercial use. Day-one runtime support covers llama.cpp, MLX, vLLM, SGLang, ONNX, and Liquid LEAP.
How does LFM2.5-8B-A1B compare to dense 8B models?
On the same hardware, LFM2.5-8B-A1B generates tokens 3-5x faster than a dense 8B like Llama 3.1 8B because only 1.5B parameters fire per token. It also ships with a 128K context window and an explicit reasoning trace before its final answer. Dense 8B models still tend to win on raw quality at single-turn tasks, but LFM2.5-8B-A1B is the better pick for low-latency on-device use, long-context tasks, and multilingual workloads.
Check VRAM for LFM2.5-8B-A1B, or compare hardware options.
Related Guides
Best GPU for Running LLMs Locally
Top picks across every budget tier in 2026.
How Much VRAM Do I Need?
Match your GPU VRAM to the largest model you can run.
LLM Quantization Guide
Q4 vs Q8 vs FP16 — what each tier costs in quality.
Best LLMs to Run Locally
Curated picks across model families and sizes.
LM Studio vs Ollama
Which local-LLM runtime fits your workflow.
Sources & methodology
VRAM and tokens-per-second figures are computed from the published spec using the sitewide formula on the methodology page. Primary sources for this guide:
- Hugging Face model card. Parameter count, license, runtime support.
- LFM2.5-24B-A2B and GGUF. 24.0B total, 2.45B active, llama-cli command, file sizes checked 13 October 2026.
- Liquid AI release blog. 8.3B total / 1.5B active, 128K context, 18.5K tok/s on H100, 253 tok/s on M5 Max.
- LiquidAI/d1-3B and d1-3B-GGUF. 3.12B total, llama-server command, file sizes checked 12 October 2026.
- Open d1 blog, 7 October 2026. Decision models, not chat models.
Spot a number that does not match the linked source? Contact us and the guide will be updated.