Tri 21B Think — Hardware Requirements & GPU Compatibility
ChatReasoning
Tri 21B Think is a 20.7B-parameter open language model from trillionlabs. It supports a context window of up to 32,768 tokens. At BF16 it needs about 42.17 GB of VRAM — see which GPUs and Macs can run it below.
Tri 21B Think runs on a 48 GB+ Mac or a workstation GPU
Best local pick → Mac Studio M5 Max (48 GB) · BF16 · 45.1 GB · ~10 tok/s
Macs from 48 GB (BF16 · ~10 tok/s)
Apache 2.0
Not set up for it?
Reference GPU · RTX 4090 24 GB
Estimates, ±20% · how we estimate
Run Tri 21B Think
One recommended way. Every other method, OS and quant is in the step-by-step guide.
Recommended
vllm serve trillionlabs/Tri-21B-Think --max-model-len 16384vLLM repo with its own quantisation
Download: Weights on Hugging Face
Other ways
each opens step-by-stepRun in cloud
Fits on A100 80GB (34 GB headroom) · BF16
- Generation speed
- ~31 tok/s
- generation speed
- Cost per 1M output tokens
- $10.04
- per 1M output tokens
How Much VRAM Does Tri 21B Think Need?
Select a quantization level; the listed hardware adheres to it.
| Quantization | VRAM | +32K | On RTX 4090 (reference) | |
|---|---|---|---|---|
| 42.2 GB | 48.5 GB | 16-bit · bfloat — Brain floating point 16 — preferred for training | Does not fit: needs 45.1 GB |
Selected BF16
Brain floating point 1645.1GBat 16K context
- Weights
- 41.5 GB
- KV cache
- 3.4 GB
- Overhead
- 0.3 GB
Doesn't fit RTX 4090 (reference) · needs 21.1 GB more
Which Hardware Can Run Tri 21B Think?
BF16 · 42.2 GBTri 21B Think (BF16) requires 42.2 GB of VRAM to load the model weights. For comfortable inference with headroom for KV cache and system overhead, 55+ GB is recommended. Using the full 33K context window can add up to 6.3 GB, bringing total usage to 48.5 GB. No consumer GPU has enough memory. 67 devices with unified memory can run Tri 21B Think, including Mac Studio (M3 Ultra, 256GB), Mac Studio (M3 Ultra, 512GB), Mac Pro M2 Ultra (64 GB).
48 GB
48 GB tier →fits · 2.9 GB free · up to 16K ctx
2 Mac classes · 1 device family
M5 Max48 GB~10Simulate output on Mac Studio M5 Max (48 GB)M5 Pro48 GB~5Simulate output on Mac Mini M5 Pro (48 GB)Snapdragon X2 Elite Extreme (X2E-96-100)~4Simulate output on Snapdragon X2 Elite Extreme Copilot+ PC
By VRAM tier · typical GPUs and Macs at BF16
Similar models
4| Model | Score | VRAM | tok/s | Why it is similar |
|---|---|---|---|---|
| InternVL3 5 GPT OSS 20B A4B Preview HFOpenGVLab | — | 13.6 GB | ~180 | same size |
| Internlm 20BInternLM | — | 45.3 GB | —needs > 24 GB | same size |
| ERNIE 4.5 21B A3B ThinkingBaidu | — | 14.3 GB | ~213 | same size |
| Mistral Small Instruct 2409Mistral AI | — | 17.4 GB | ~46 | same size |
Specifications
- Publisher
- trillionlabs
- Parameters
- 20.7B
- Architecture
- TrillionForCausalLM
- Context Length
- 32,768 tokens
- Vocabulary Size
- 124,416
- Release Date
- 2026-02-18
- License
- Apache 2.0
Frequently Asked Questions
How much VRAM does Tri 21B Think need?
Tri 21B Think requires 42.2 GB of VRAM at BF16. Full 33K context adds up to 6.3 GB (48.5 GB total).
VRAM = Weights + KV Cache + Overhead
Weights = 20.7B × 16 bits ÷ 8 = 41.5 GB
KV Cache + Overhead ≈ 0.7 GB (at 2K context + ~0.3 GB framework)
Fit ratings and hardware model lists check this model with room for a 16K-token context, which needs a little more memory.
KV Cache + Overhead ≈ 7 GB (at full 33K context)
VRAM usage by quantization
Can NVIDIA GeForce RTX 5090 run Tri 21B Think?
No — Tri 21B Think requires at least 42.2 GB at BF16, which exceeds the NVIDIA GeForce RTX 5090's 32 GB of VRAM.
Can I run Tri 21B Think on a Mac?
Yes — Mac Studio M5 Max (48 GB) and 55 other Macs can run Tri 21B Think. Apple Silicon uses unified memory, so the model shares RAM with the system. At BF16 you need at least 42.2 GB of usable unified memory (RAM minus macOS overhead).
Can I run Tri 21B Think locally?
Yes — Tri 21B Think can run locally on consumer hardware. Popular tools include Ollama, LM Studio, and llama.cpp.
How fast is Tri 21B Think?
At BF16, Tri 21B Think can reach ~114 tok/s on AMD Instinct MI350X. Speed depends mainly on GPU memory bandwidth. Real-world results typically within ±20%.
tok/s = (bandwidth GB/s ÷ model GB) × efficiency
Example: NVIDIA B200 → 8000 ÷ 42.2 × 0.65 = ~123 tok/s
Estimated speed at BF16 (42.2 GB)
Real-world results typically within ±20%. Speed depends on batch size, quantization kernel, and software stack.
What's the download size of Tri 21B Think?
At BF16, the download is about 41.45 GB.
Which GPUs can run Tri 21B Think?
No single consumer GPU has enough VRAM to run Tri 21B Think at BF16 (42.2 GB). Multi-GPU or professional hardware is required.
Which devices can run Tri 21B Think?
67 devices with unified memory can run Tri 21B Think at BF16 (42.2 GB), including ASUS Ascent GX10, Beelink GTR9 Pro (Ryzen AI Max+ 395, 128 GB), Framework Desktop (Ryzen AI Max+ 395, 128 GB), GMKtec EVO-X2 (Ryzen AI Max+ 395, 128 GB). Apple Silicon Macs use unified memory shared between CPU and GPU, making them well-suited for local LLM inference.