Unitree H1-2

The Unitree H1-2 is a 178 cm, 70 kg full-size humanoid robot built for researchers and engineers who need a bipedal platform that can carry compute and sensors instead of putting them on a cart. Unitree ships it with an Intel i5/i7 box inside the torso and gives you three bays for Jetson Orin NX modules—up to 3× 100 TOPS each—so you can co-locate perception, planning, and LLM inference on the robot instead of tethering to a workstation. That makes the H1-2 one of the few humanoids under six figures that can run a 70 B parameter model locally while walking around.

Unlike toy-size humanoids, the H1-2’s 360 N·m knee joints and 120 N·m shoulder actuators let it haul a 7 kg continuous payload (21 kg peak)—enough for a UPS, extra batteries, or a second GPU module. The 864 Wh swappable battery keeps the robot alive for ~2 h, so you get a full working window before the next hot-swap. If your use case is “put an AI agent inside a body that can open doors, carry tools, and answer questions,” this is the cheapest way to do it without writing a DARPA grant.

• Base CPU: Intel Core i5 or i7 (12-gen U-series, 15 W default)
• AI bays: 3× M.2-2280 slots pre-wired for Jetson Orin NX 8 GB or 16 GB modules
• 1× Orin NX 16 GB = 100 TOPS INT8, 64 GB/s LPDDR5
• 3× Orin NX 16 GB = 300 TOPS, 192 GB/s aggregate—comparable to a desktop RTX 4090 but at 45 W total GPU power

• Jetson Orin NX 16 GB: fits 70 B models at 4-bit, 33 B at 8-bit, 8 B at FP16
• Dual Orin NX 16 GB: split-model pipeline (embeddings on GPU-0, generation on GPU-1) keeps context <4 GB per GPU, so 128 k tokens is feasible
• CPU RAM: 32 GB DDR4 soldered; 64 GB SKU available on request—enough to cache LoRA adapters or run a second distilled model

• Idle robot body: 120 W
• Walk + generate 30 tokens/s: ~250 W total; battery life drops to 1 h at that load
• Fan noise: 42 dB(A) at 1 m—quiet enough for office demos

• Prompt: 1,024 tokens → 65 tokens/s
• Generation: 45–50 tokens/s sustained at 25 W GPU
• Long-context 128 k: 28 tokens/s, context-switch latency 380 ms

• Llama 3.1 8B Q4_K_M: fits in 8 GB VRAM, >45 tokens/s—perfect for real-time voice chat
• DeepSeek-R1-Distill-Llama-70B Q2_K: 36 GB, needs dual Orin NX 16 GB; runs at 18 tokens/s, reasoning quality within 3 % of full-precision
• Qwen2.5-14B Q6_K: 11 GB, single Orin NX 16 GB, 38 tokens/s—strong multilingual agent
• Mixtral-8×7B Q3_K_M: 24 GB, split across two Jetsons with tensor parallelism, 22 tokens/s

• Llava-1.6-34B Q4: 19 GB, works on dual Orin; 15 tokens/s including 336×336 vision encoder
• Optional wrist RealSense: RGB-D stream fused with LiDAR point-cloud; Jetson encoder handles 30 fps 1280×720 with <80 ms end-to-end latency

• 4-bit: almost no perceptible quality drop on 8–30 B models; 30 % higher tokens/s vs 5-bit
• 3-bit: acceptable for 70 B reasoning models; 2-bit starts to hallucinate on math tasks

Drop a 70 B model on the robot, give it a semantic map of the building, and you have a walking conversational guide that remembers prior interactions—no cloud calls, no GDPR headaches.

Oil rigs and telecom towers have poor LTE. Run a vision-language fault detector (Qwen-VL-Chat 9B) locally; the robot climbs stairs, snaps photos, annotates defects, and logs JSON over Wi-Fi 6 when back in range.

At ~$112 k the H1-2 is still serious money, but it’s 4× cheaper than Boston Dynamics Stretch and 2× cheaper than 1X Neo beta. If you want a human-scale body for your open-source agent, this is the entry ticket.

Three Orin NX modules give you 300 TOPS in a 70 kg roaming package. Put one robot in each warehouse aisle; it serves quantized 33 B models to handheld scanners via Ethernet—no fixed rack required.

• G1 tops out at one Orin Nano 8 GB—good for 7 B models only
• H1-2 carries 3× Orin NX 16 GB, so 70 B is feasible
• G1 is $30 k; H1-2 is $112 k—pay for payload and compute headroom

• Phoenix ships with Nvidia A6000 (48 GB) but is tethered for power
• H1-2 runs untethered, swappable battery, 360° LiDAR included
• Phoenix torso is 17 kg lighter; H1-2 arm torque is 50 % higher—pick your trade-off

Pick the H1-2 when you need a self-contained, walking inference platform that can host 30–70 B parameter models without external GPUs or cloud APIs. Pick something lighter (G1) or stationary (Phoenix) if you don’t need the full 7 kg payload or 2-hour untethered runtime.