exla E1 inside · 192 GB · 2.05 TB/s
smol.
A 300-billion-parameter model that lives on your desk. Hundreds of tokens a second. Nothing leaves the room.
scroll
01 · speed
Faster than the fastest Mac. Same model.
Decoding a giant model is a memory problem, not a compute problem. smol puts 192 GB right against the chip at 2 TB/s, so every token arrives before you finish reading the last one.
DeepSeek-V4-Flash, single user, 32K context. smol figure is from the exla E1 cycle-level model (silicon in development). Mac figure is its memory-bandwidth ceiling at the same weight format; measured today it is lower.
02 · inside
One block of aluminium. Everything it needs, nothing else.
- 1,089 holes2.6 mm on a 4.4 mm grid. The whole top is the exhaust.
- 120 mm fan at 1,100 rpmAbout 21 dBA. You will not hear it.
- Copper vapor chamber, 50 finsKeeps the chip near 68 °C on a warm desk.
- exla E1Three 4 nm chiplets and twelve LPDDR5X stacks under one lid. 192 GB.
- Wi-Fi 7, 2.5 GbE, USB-CAntennas sit in the side slot, the only opening the radio can use.
- 2 TB SSD underneathHolds six models. Swaps one into memory in about 25 seconds.
03 · how
Every token reads only the experts it needs.
Today's best open models are mixtures of experts: hundreds of small networks, of which a router wakes eight per token. E1 streams just those weights from memory, decodes them from a 4-bit learned codebook on the fly, and splits every matrix across three chiplets so all 1,536 bits of memory bus work at once.
- bus
- 1,536-bit LPDDR5X
- read per token
- 7.05 GB
- experts awake
- 8 of 256
- weights
- 4.25 bits / weight
04 · setup
Plug it in. That's the setup.
One figure-8 cord. No brick. The display wakes and shows the model.
Cable, and it is on your network. No cable, and it shows a setup Wi-Fi to join from your phone.
Open it in any browser, or point any OpenAI or Anthropic app at it. Done.
05 · build
Speaks the APIs you already use.
OpenAI and Anthropic compatible, so Claude Code, Cursor, Open WebUI and every SDK work by changing one URL. Or use the smol SDK: zero config, typed, and an emulator so you can build before your box arrives.
06 · spec
smol, rev A
- size
- 180 × 180 × 64 mm 1.7 kg · one-piece 6061 aluminium
- chip
- exla E1 3 × TSMC N4P chiplets
- memory
- 192 GB LPDDR5X-10667 · 2.05 TB/s
- storage
- 2 TB NVMe about six models
- speed
- 194–440 tok/s GLM-5.3 · DeepSeek-V4 · Qwen3.8
- power
- ~129 W ~23 W idle · 180 W internal supply
- noise
- ~21 dBA one 120 mm fan
- connect
- 2.5 GbE · Wi-Fi 7 · USB-C OpenAI + Anthropic APIs
smol is in development. Speeds and power are design targets from the exla E1 cycle-level model and thermal budget, not measurements of shipping hardware.