FENRIS COMPUTERS

FENRIS COMPUTERS — STOCKHOLM, SWEDEN

Open-source AI, running where it has no business running.

Swedish engineering applied to inference. We take open weights and old silicon and make them behave like something newer, faster, and smarter than either was meant to be.

SEE THE BENCHMARKS   KEEP ME UPDATED

WHAT WE DO

  • OPTIMIZEOpen-source models, compressed and rebalanced for hardware they were never designed for.
  • TUNEKernels, quantization, and memory layout, tuned per device, not per spec sheet.
  • ORCHESTRATEOld hardware plus a few new components, scheduled as one coherent inference cluster.
  • ENHANCEAn inference engine built from the ground up. Faster tokens, sharper answers, same silicon.

BENCHMARKS — QWEN3.8-27B

Official scores at full precision, plus what actually survives quantization. More models land as we finish tuning them.

MODEL PARAMS QUANT BENCHMARK SCORE SOURCE
Qwen3.8-27B 27B BF16 GPQA Diamond 89.2 [OFFICIAL] model card
Qwen3.8-27B 27B BF16 LiveCodeBench v6 90.3 [OFFICIAL] model card
Qwen3.8-27B 27B BF16 SWE-bench Pro 61.7 [OFFICIAL] model card
Qwen3.8-27B 27B BF16 IFBench 79.5 [OFFICIAL] model card
Qwen3.8-27B 27B FP8 SWE-bench Pro 61.7 [OFFICIAL] model card
Qwen3.8-27B 27B FP8 LiveCodeBench v6 90.3 [OFFICIAL] model card
Qwen3.8-27B 27B GGUF Q8_0 GPQA Diamond ~89 [REPRODUCED] Quesma
Qwen3.8-27B 27B GGUF Q4_K_M GPQA Diamond ~88 [REPRODUCED] Quesma
Qwen3.8-27B 27B GGUF Q4_K_M Terminal-Bench 2.1 ~75 [REPRODUCED] Quesma
Qwen3.8-27B 27B GGUF UD-Q2_K_XL GPQA Diamond ~86 [REPRODUCED] Quesma
Qwen3.8-27B 27B GGUF UD-IQ1_S GPQA Diamond ~50 (collapse) [REPRODUCED] Quesma
Qwen3.8-2.4T-A95B 2.4T-A95B BF16 GPQA Diamond 92.6 [OFFICIAL] model card

[OFFICIAL] = reported by the model author   [REPRODUCED] = independently verified, source linked

KEEP ME UPDATED

New benchmarks, new hardware, new models we've tuned. Nothing else. No spam.