Mellum2.1 12B Thinking
MoE consumerA code model with no chat layer. JetBrains shipped Mellum2.1 Thinking on October 8, 2026: a 12.15B-total, 2.5B-active Apache 2.0 model trained for the code workflow the company’s IDEs are built around, and the official GGUF landed in the same week. 7,800 downloads on the GGUF build within a day.
What one-workflow training buys. A general 12B chat model spends its capacity on general text; Mellum’s 2.5B active slice is tuned to find, read, and complete code across a project, at the 131k context its config carries. About 7.0GB in 4-bit, it targets the same box a 16GB machine already hosts, with a 56MB-per-1,000-token cache that fits comfortably in laptop memory budgets.
Where it sits. Jetbrains’ first Mellum (April 2026, 4B) proved the focused-specialist idea; 2026’s 12B line adds the thinking variant for harder retrievals. Against qwen3 coder-line generalists it trades breadth for latency-per-answer, which is the trade an IDE needs.
- 12.2B
- 131k
- apache 2.0
- Oct 2026
Save your hardware and every model page answers the real question: will it run on your machine, and how fast?
Join free - save your rig →Run it locally
Per-quant memory needs and a static "can you run it?" reference - no rig entry required
The reference hardware
22 reference configs, drawn in-house. Scroll for more.
Can you run it? - reference rigs
| Rig | Q4_K_M | FP16 |
|---|---|---|
| Single GTX 1080 Ti (11GB) | tight | no -> cloud |
| 4x H100 80GB (320GB) | fast 4428.2t/s | fast 1410.8t/s |
| NVIDIA DGX Station 748GB | fast 2643.7t/s | fast 842.3t/s |
| 8x RTX 3090 rack (192GB) | fast 2475.0t/s | fast 788.5t/s |
| 4x RTX 5090 (128GB) | fast 2368.8t/s | fast 754.7t/s |
| AMD Instinct MI300X (192GB) | fast 1759.7t/s | fast 560.6t/s |
| 4x RTX 4090 (96GB) | fast 1332.4t/s | fast 424.5t/s |
| 2x RTX 5090 (64GB) | fast 1184.4t/s | fast 377.3t/s |
| 2x RTX 3090 (48GB) | fast 618.8t/s | fast 197.1t/s |
| Single RTX 5090 (32GB) | fast 592.2t/s | fast 188.7t/s |
| RTX PRO 6000 Blackwell (96GB) | fast 592.2t/s | fast 188.7t/s |
| Mac Studio M4 Ultra 192GB | fast 393.7t/s | fast 125.4t/s |
| Mac Studio M4 Ultra 512GB | fast 393.7t/s | fast 125.4t/s |
| Single RTX 4090 (24GB) | fast 333.1t/s | offload |
| MacBook Pro M5 Max 128GB | fast 221.4t/s | fast 70.5t/s |
| Dual EPYC 9004 + 768GB DDR5-4800 | fast 152.3t/s | fast 48.5t/s |
| DGX Spark 128GB unified | fast 90.2t/s | fast 28.7t/s |
| Ryzen AI Max+ 395 128GB | fast 84.6t/s | fast 27.0t/s |
| Jetson AGX Orin 64GB | fast 67.7t/s | fast 21.6t/s |
| Epyc + 512GB DDR4-3200 + 2x RTX 3090 | fast 67.7t/s | fast 21.6t/s |
| Epyc + 512GB DDR4-2400 + 2x RTX 3090 | fast 50.8t/s | ok 16.2t/s |
| NVIDIA Jetson Orin NX 16GB | fast 33.8t/s | no -> cloud |
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.
Download options
Or run it in the cloud
No per-token API provider pricing tracked for Mellum2.1 12B Thinking yet. For flagship list prices, see the calculator.
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.