Liquid’s edge multimodal, 3B of it. d1-3B (October 7, 2026) is a LiquidAI edge model with the official GGUF out the same day: image and text in and out on a phone-class footprint. The vision stack sits on the company’s LFM2-VL architecture (30 layers, 8 KV heads x 64 head-dim, 32,768 context); about 1.8GB in 4-bit and a 60MB-per-1,000-token cache.
The license is bespoke. Liquid ships its edge line under its own license rather than a standard open one, terms worth reading before any product embedding; the weights are downloadable and the day-one GGUF keeps the deployment path short.
9,088 downloads and 273 likes in the first days.
multimodal
edge
vision
- Parameters
-
3.1B
- Context
- 32k
- License
- liquid ai license
- Origin
-
🇺🇸
USA
- Released
- Oct 2026
Run it locally
Per-quant memory needs and a static "can you run it?" reference - no rig entry required
Q4_K_M
1.8GB
4.0GB min
8.0GB rec
Balanced - the usual local sweet spot
The reference hardware
22 reference configs, drawn in-house. Scroll for more.
Can you run it? - reference rigs
| Rig |
Q4_K_M |
| 4x H100 80GB (320GB) |
fast 3612.8t/s
|
| NVIDIA DGX Station 748GB |
fast 2156.9t/s
|
| 8x RTX 3090 rack (192GB) |
fast 2019.3t/s
|
| 4x RTX 5090 (128GB) |
fast 1932.6t/s
|
| AMD Instinct MI300X (192GB) |
fast 1435.6t/s
|
| 4x RTX 4090 (96GB) |
fast 1087.1t/s
|
| 2x RTX 5090 (64GB) |
fast 966.3t/s
|
| 2x RTX 3090 (48GB) |
fast 504.8t/s
|
| Single RTX 5090 (32GB) |
fast 483.1t/s
|
| RTX PRO 6000 Blackwell (96GB) |
fast 483.1t/s
|
| Mac Studio M4 Ultra 192GB |
fast 321.2t/s
|
| Mac Studio M4 Ultra 512GB |
fast 321.2t/s
|
| Single RTX 4090 (24GB) |
fast 271.8t/s
|
| MacBook Pro M5 Max 128GB |
fast 180.6t/s
|
| Single GTX 1080 Ti (11GB) |
fast 130.5t/s
|
| Dual EPYC 9004 + 768GB DDR5-4800 |
fast 124.2t/s
|
| DGX Spark 128GB unified |
fast 73.6t/s
|
| Ryzen AI Max+ 395 128GB |
fast 69.0t/s
|
| Jetson AGX Orin 64GB |
fast 55.2t/s
|
| Epyc + 512GB DDR4-3200 + 2x RTX 3090 |
fast 55.2t/s
|
| Epyc + 512GB DDR4-2400 + 2x RTX 3090 |
fast 41.4t/s
|
| NVIDIA Jetson Orin NX 16GB |
fast 27.6t/s
|
Fit tiers use the same will-it-run logic as the rig finder. For comfortable fits, the badge reflects decode speed: fast >=20 t/s, ok 8-20 t/s, slow <8 t/s. t/s is a bandwidth estimate, not a measured benchmark.
PRICE HISTORY
Inference cost over time
Data accumulates from the first daily sync - longer ranges populate over time. Prices come from OpenRouter snapshots, not a historical API.
Loading price history...