Leaderboard / Trillion Labs
Tri-21B Think
Trillion Labs · Tri · released 2026-02-18 · data checked 2026-11-01
Tri-21B Think is a 20.7B-parameter dense model from Trillion Labs (Rest of Asia), released 2026-02-18 under the Apache-2.0 license. It scores 43.6 on the Small LM Index, #30 of 34 medium models. It needs about 15.1 GB of memory at 4-bit with an 8K-token context.
Small LM Index
43.6
#30 of 34 scored medium models
Likely 39.5–47.7 · from 5 of 6 core benchmarks
Parameters
20.7B
Dense
Price per 1M tokens
—
No hosted API on OpenRouter; self-host
Context window
32K
tokens
Benchmarks
Each bar shows this model's score. The tick marks the best medium model on that benchmark. Sources: 8 reported by Trillion Labs. How we label sources.
Not reported: estimated
What the index model expects from this model's other scores. Not a measurement.
SWE-bench Verified≈ 24
Specifications
- Lab
- Trillion Labs (KR · Rest of Asia)
- Architecture
- Dense transformer
- Total parameters
- 20.7B
- Active parameters
- All (dense)
- Size class
- Medium
- Context window
- 32K tokens
- Input modalities
- text
- Reasoning
- Reasoning (thinking)
- License
- Apache-2.0
- Openness
- Open weights
- Memory to run
- ~15.1 GB at 4-bit, 8K context (more sizes)
- Hugging Face
- 459 downloads (30 days) · 29 likes
Compare with
- Tri-21B Think vs gpt-oss-20b (20.9B, index 64.0)
- Tri-21B Think vs EuroLLM 22B Instruct (22.6B, index 19.2)
- Tri-21B Think vs LFM2.5 24B-A2B (24B, index 61.4)
- Tri-21B Think vs Magistral Small 1.2 (24B, index 55.9)
Sources
Something wrong here? Report it.
Run it locally
No GGUF build from a well-known publisher yet.
vLLM (GPU server)
vllm serve trillionlabs/Tri-21B-ThinkFull-precision weights, for a server GPU.
Memory needed: weights, context cache and runtime buffers.
| Quantization | 4K context | 32K context |
|---|---|---|
| 16-bit F16 | 43 GB | 49 GB |
| 8-bit Q8_0 | 24 GB | 29 GB |
| 4-bit Q4_K_M | 14 GB | 20 GB |