LiquidAI · embedding
LFM2.5-ColBERT-350M
LiquidAI/LFM2.5-ColBERT-350MLFM2.5-ColBERT-350M at Q4_K_M is exactly 228,407,552 bytes (0.21 GiB / 0.23 GB) — an effective 5.172 bits per weight, not the nominal 4.
From the file· summed from 1 file(s)From the file· KV per layer
Parameters
353M
Architecture
lfm2
16 layers
Context
128,000
native (config.json)
License
other
Shipped quantizations
● exact bytes, summed from published files
| Quant | Size● | Exact bytes● | Effective bpw● | Tensors● | Publisher |
|---|---|---|---|---|---|
| Q2_K | 0.15 GiB | 159,662,784 | 3.615 | — | mradermacher |
| Q3_K_S | 0.17 GiB | 180,229,824 | 4.081 | — | mradermacher |
| Q3_K_M | 0.18 GiB | 192,231,104 | 4.353 | — | mradermacher |
| Q3_K_L | 0.19 GiB | 202,127,040 | 4.577 | — | mradermacher |
| IQ4_XS | 0.20 GiB | 210,897,600 | 4.775 | — | mradermacher |
| Q4_0 | 0.20 GiB | 218,405,120 | 4.945 | — | LiquidAI |
| Q4_K_S | 0.20 GiB | 219,847,360 | 4.978 | — | mradermacher |
| IQ4_XS | 0.21 GiB | 225,004,640 | 5.095 | — | cstr |
| Q4_K_M | 0.21 GiB | 228,407,552 | 5.172 | — | LiquidAI |
| Q4_K_M | 0.21 GiB | 228,408,000 | 5.172 | — | mradermacher |
| Q4_K | 0.22 GiB | 233,987,168 | 5.298 | — | cstr |
| Q5_K_S | 0.24 GiB | 254,335,680 | 5.759 | — | mradermacher |
| Q5_K_M | 0.24 GiB | 259,488,000 | 5.875 | — | LiquidAI |
| Q5_K_M | 0.24 GiB | 259,488,448 | 5.875 | — | mradermacher |
| Q5_K | 0.25 GiB | 269,917,280 | 6.112 | — | cstr |
| Q6_K | 0.27 GiB | 292,510,976 | 6.623 | — | LiquidAI |
| Q6_K | 0.27 GiB | 292,511,424 | 6.623 | — | mradermacher |
| Q8_0 | 0.35 GiB | 377,707,616 | 8.552 | — | cstr |
| Q8_0 | 0.35 GiB | 378,097,248 | 8.561 | — | LiquidAI |
| Q8_0 | 0.35 GiB | 378,097,696 | 8.561 | — | mradermacher |
| F32 | 0.66 GiB | 709,009,312 | 16.053 | — | cstr |
| BF16 | 0.66 GiB | 709,399,008 | 16.062 | — | LiquidAI |
| F16 | 0.66 GiB | 709,399,008 | 16.062 | — | LiquidAI |
| F16 | 0.66 GiB | 709,399,456 | 16.062 | — | mradermacher |
No KV cache
architectural, not a gap in our data
This architecture allocates no KV cache. Encoder and embedding models process their input in one pass rather than generating token by token, so there is nothing to carry forward between steps and memory does not grow with context. Its footprint is the weights plus a working buffer, and that is the whole story.
Compare with
same modality, comparable size
Will it run on your card?
full quant x context sweep
Radeon RX 6500 XT 4GBGeForce RTX 3050 6GBGeForce RTX 5050 8GBGeForce RTX 3080 10GBGeForce RTX 2080 Ti 11GBGeForce RTX 5070 12GBGeForce RTX 5060 Ti 16GBApple M3 Pro 18GBGeForce RTX 3080 Ti 20GBGeForce RTX 5090 D V2 24GBGeForce RTX 5090 D 32GBApple M5 Max 36GBApple M5 Pro 48GBApple M5 Max 64GBApple M3 Ultra 96GBApple M5 Max 128GBApple M2 Ultra 192GBApple M3 Ultra 256GBApple M3 Ultra 512GB
Why other calculators give a different number
A parameters × bits ÷ 8 estimate puts Q4_K_M at roughly 0.19 GiB. The real file is 0.21 GiB, because a quantization is a mixture and some tensors are always kept at higher precision.
Architecture
from config.json
Layers
16
Attention heads
16
KV heads
8
Head dim
64
Hidden size
1024
Vocab
64,402
Sliding window
none
SWA period
—
MLA
no
Experts
—
Experts per token
—
use_sliding_window
—
Questions people ask
- How much VRAM does LFM2.5-ColBERT-350M need?
- Q4_K_M is exactly 228,407,552 bytes (0.21 GiB) in weights. Add the KV cache, which depends on your context length, plus roughly half a gigabyte of runtime overhead.
- Which quantization of LFM2.5-ColBERT-350M should I use?
- Q4_K_M is the usual default. Pick the largest quantization that fits your card at the context you actually need — the table above gives exact sizes for every one published.