Runs on ≤12 GB · Hugging Face ↗
Not independently scored by Artificial Analysis
JetBrains' purpose-built coding MoE — 12B total, ~2.5B active, trained on ~10.6T tokens for code completion and infill. The rare coding specialist that actually fits a 16 GB card, and ~2x faster than dense models its size. Custom `MellumForCausalLM` arch.
BigCodeBench-Hard pass@1 27% (40/148), via official protocol · vLLM · one request at a time · a shared cluster GPU · mean of 3 runs, range 0.7 pp. The brutal counterpart to HumanEval — where HumanEval saturates near the top, BCB-Hard spreads the field, so this is the number that actually separates coding ability. served artifact not verified; different runtime
Our own runs on this model. Each number carries how many problems it is over and how many times we ran it — these are not comparable with each other, and none of them is on the same scale as the third-party indices on the leaderboard.
| test | score | over | conditions |
|---|---|---|---|
| BigCodeBench-Hard | 27.3% | n=148 · 3 run(s) | official protocol · vLLM · one request at a time · a shared cluster GPU · mean of 3 runs, range 0.7 pp |
Real GGUF file sizes = weight VRAM. Add the KV cache for your context (≈1 GB at 32K, fp16). Read from the GGUF header, but this architecture has never been checked against llama.cpp — treat it as an estimate.
| Quantization | Size |
|---|---|
| Q2_K | 5.0 GB |
| Q3_K_S | 5.8 GB |
| Q3_K_M | 6.3 GB |
| Q3_K_L | 6.6 GB |
| IQ4_XS | 6.7 GB |
| Q4_K_S | 7.4 GB |
| Q4_K_M | 8.1 GB |
| Q5_K_S | 8.6 GB |
| Q5_K_M | 9.2 GB |
| Q6_K | 10.9 GB |
| Q8_0 | 12.9 GB |
~8 GB at Q4 — fits an 8-16 GB card. Built for fast completion / fill-in-the-middle rather than long agentic chains, so read a single-shot BCB accordingly. Trending in every "best local coding model" roundup this week.