Runs on ≤32 GB · Hugging Face ↗
Not independently scored by Artificial Analysis
Released 2026-09-16 by XingChen AGI, Apache-2.0. A mixture-of-experts: 31.2B parameters in total, 64 experts with 4 active per token, 40 layers and a 256K context, on its own Xing4_0ForCausalLM architecture.
BigCodeBench-Hard pass@1 19% (28/148), via official protocol · llama.cpp IQ4_NL · one request at a time · a shared cluster GPU · mean of 3 runs, range 0.0 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.
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 | 18.9% | n=148 · 3 run(s) | official protocol · llama.cpp IQ4_NL · one request at a time · a shared cluster GPU · mean of 3 runs, range 0. |
| BFCL v4 non-live | 65.2% ± 3.0 | n=1390 · 1 run(s) | BFCL v4 · IQ4_NL GGUF via llama.cpp · one request at a time · a shared cluster GPU |
| RULER (long context) | 88.3 ± 4.7 | N=5/task · at 32K | 1 quantisations measured |
65.2% ± 3.0 over 1390 problems
IQ4_NL · a shared cluster GPU · KV f16 · 1 run. This is the non-agentic half of BFCL: does the model call the right function with the right arguments. It says nothing about how it behaves over a long agentic conversation, which is a separate measurement. The score is BFCL’s own: the unweighted mean of four groups (simple, multiple, parallel, parallel multiple), where simple is itself the mean of Python, Java and JavaScript. Every group counts the same whatever its size, and irrelevance detection — which we also measure, another 240 problems — is not part of it. The margin follows that same weighting.
| quant | GiB | 4K | 8K | 32K |
|---|---|---|---|---|
| IQ4_NL | 7.3 | 86.0 ± 5.9 · N=5 | 86.5 ± 6.4 · N=5 | 88.3 ± 4.7 · N=5 |
All 13 RULER tasks, one request at a time, KV cache f16, bench commit c3f5e3b4, seed 42. The score is the mean of the 13; the margin is two standard deviations of the sampling, so it grows as the sample shrinks and as the score drops. Everything here clears the paper’s effective-length bar of 85.6 at every length shown.
This is not a coding score and must not be read as one. RULER asks whether the model finds and follows a fact buried in a long text. Quantising hits the two differently, so a model that still retrieves at length may already have lost the code.
Real GGUF file sizes = weight VRAM. Add the KV cache for your context (≈19 GB at 32K, fp16). Estimated from the original model config, not from the GGUF we serve. Less reliable than the rest of this page.
| Quantization | Size |
|---|---|
| IQ4_NL | 20.1 GB |
Upstream llama.cpp does not load it yet: we built pull request #29012, which was not merged when we measured. The vendor publishes a single IQ4_NL build, so that is the only point we have. Artificial Analysis does not score it.