niklaslenzHome / Game
Game test
The Gauntlet biathlon
Every setup plays one solo race in a biathlon game. The model runs a course with four shooting ranges. At each range it gets one task and one answer, with no feedback and no retry. Wrong parts cost penalty loops. The finish time sets the rank in the memory class. How it works.
season seed 2148 finished, verifiedupdated Nov 11, 2026
42 of 48 races ran with thinking off. In 6 of them the model's thinking can't be switched off; they are marked. These runs show speed under pressure, not what the models can do with reasoning. Thinking-on races will get their own list.
16 GB · RTX 5060 Ti 16 GB
16 GB class
| # | Setup | Finish time | Targets hit | Penalty loops | Thinking |
|---|---|---|---|---|---|
| 1 | Qwen3.8-27B EXL3 3.0bpwRTX 5060 Ti 16 GB · Mia's one-click EXL3 kit (Qwen3.8-27B 3.0 bpw)Thinking off | 256.5 s | 9 of 20 | 11 | 28.7 s |
| 2 | Qwen3.8-27B ByteShapeRTX 5060 Ti 16 GB · llama.cpp b12223, ByteShape Qwen3.8-27B IQ4_XS GGUF with MTP draftsThinking off | 262.0 s | 9 of 20 | 11 | 34.2 s |
| 3 | Qwen3.8-27B EXL3 2.5bpwRTX 5060 Ti 16 GB · Mia's one-click EXL3 kit (Qwen3.8-27B 2.5 bpw, MTP drafts)Thinking off | 278.2 s | 8 of 20 | 12 | 38.4 s |
| 4 | Qwen3.8-27B TQ_XSRTX 5060 Ti 16 GB · tqLLM 1.5.0 (Docker), Qwen3.8-27B TQ_XS + DFlash2-TQ3 drafts, TQ3 KVThinking off | 283.5 s | 7 of 20 | 13 | 22.7 s |
| 5 | Clover 35BRTX 5060 Ti 16 GB · llama.cpp (Clover build), Diffuspace Clover-35B-A3B-Preview (Q4_0 GGUF)Thinking off | 326.5 s | 11 of 20 | 9 | 131.7 s |
| 6 | Qwen3.8-35B-A3B-DistillRTX 5060 Ti 16 GB · llama.cpp b12223, Q4_K_M GGUF (MoE partly on CPU)Thinking off | 347.1 s | 5 of 20 | 15 | 68.3 s |
| 7 | Raon-1-miniRTX 5060 Ti 16 GB · llama.cpp, KRAFTON Raon-1-mini-preview (IQ4_XS GGUF, DFlash2)Thinking off | 387.9 s | 0 of 20 | 20 | 37.1 s |
| 8 | Muse Glimmer 30BRTX 5060 Ti 16 GB · llama.cpp b11972, Unsloth UD-IQ3_XXS GGUF of Meta Muse-Glimmer-30B, with DFlash drafterThinking can't be switched off | 408.7 s | 10 of 20 | 10 | 192.9 s |
| 9 | Flint 40BRTX 5060 Ti 16 GB · llama.cpp b11972, Recursive Flint 40B (unsloth UD-Q2_K_XL GGUF)Thinking can't be switched off | 410.6 s | 5 of 20 | 15 | 134.8 s |
32 GB · RTX 5090 32 GB
32 GB class
| # | Setup | Finish time | Targets hit | Penalty loops | Thinking |
|---|---|---|---|---|---|
| 1 | ThinkingCap 27BRTX 5090 32 GB · Mia's one-click EXL3 kit (4.0 bpw, MTP drafts), 0xWhiteMage ThinkingCap-Qwen3.8-27B-KearugaThinking off | 181.6 s | 13 of 20 | 7 | 10.8 s |
| 2 | Qwen3.8-27BRTX 5090 32 GB · TensorFold 1.1.2 (BobClawblaw fork), Qwen3.8-27B MLX 4-bit, DFlash2 draftsThinking off | 188.6 s | 13 of 20 | 7 | 11.8 s |
| 3 | Qwen3.8-27B EXL3 4.0bpw MTPRTX 5090 32 GB · Mia's one-click EXL3 kit (Qwen3.8-27B 4.0 bpw, MTP drafts)Thinking off | 197.5 s | 12 of 20 | 8 | 11.7 s |
| 4 | Qwen3.8-27BRTX 5090 32 GB · vLLM 0.31.0 with humming kernels, Kaitchup Qwen3.8-27B SLQ-EAR98Thinking off | 199.6 s | 13 of 20 | 7 | 28.8 s |
| 5 | Qwen3.8-27BRTX 5090 32 GB · NInfer (Neroued), NVFP4 + MTP3Thinking off | 220.2 s | 11 of 20 | 9 | 16.4 s |
| 6 | Qwen3.6-35B-A3BRTX 5090 32 GB · TensorFold 1.2.1, Ash's recipe (NVIDIA NVFP4, DFlash2)Thinking off | 236.3 s | 10 of 20 | 10 | 20.5 s |
| 7 | Qwen3.8-27B EXL3RTX 5090 32 GB · Mia's one-click EXL3 kit (Qwen3.8-27B 4.0 bpw, DFlash2)Thinking off | 238.6 s | 9 of 20 | 11 | 10.8 s |
| 8 | Muse Glimmer 30BRTX 5090 32 GB · llama.cpp b11972, Meta Muse-Glimmer-30B GGUF (Q4_K_M 17 GB) with DFlash drafterThinking can't be switched off | 241.3 s | 13 of 20 | 7 | 70.5 s |
| 9 | Qwen3.8-27BRTX 5090 32 GB · Qwentin v0.14 (kacper-daftcode), Qwen3.8-27B EXL3 5.75 bpwThinking off | 251.9 s | 8 of 20 | 12 | 12.1 s |
| 10 | Qwen3.8-27BRTX 5090 32 GB · TensorFold 1.2.0 (Ash), BobClawblaw's qwen38-27b-vision-rtx-5090 recipe (MLX 4-bit, DFlash2)Thinking off | 256.7 s | 7 of 20 | 13 | 4.9 s |
| 11 | Nemotron 3.5 Bolt 74BRTX 5090 32 GB · TensorFold, Mia's recipe (NVIDIA Nemotron 3.5 Bolt 74B IQ2, LiLiCorr draft), 2 streamsThinking off | 275.3 s | 6 of 20 | 14 | 14.5 s |
| 12 | Nemotron 3.5 Bolt 74BRTX 5090 32 GB · TensorFold, Mia's recipe (NVIDIA Nemotron 3.5 Bolt 74B IQ2, LiLiCorr draft)Thinking off | 295.5 s | 4 of 20 | 16 | 4.7 s |
| 13 | Nanbeige4.3-33BRTX 5090 32 GB · llama.cpp b11972, Bartowski Q4_K_M GGUFThinking off | 297.1 s | 6 of 20 | 14 | 30.3 s |
| 14 | Flint 40B NVFP4RTX 5090 32 GB · llama.cpp b11803, Recursive Flint 40B (NVFP4 GGUF)Thinking can't be switched off | 354.2 s | 5 of 20 | 15 | 78.4 s |
| 15 | MiniMax M3.1 FlashRTX 5090 32 GB · Project Maya v1.1.50 (mw00), MiniMax M3.1 Flash Maya-XS (IQ1_S)Thinking off | 375.2 s | 7 of 20 | 13 | 123.4 s |
| 16 | Ripple-65BRTX 5090 32 GB · llama.cpp b12167, AtomicChat GGUF (AD-IQ3_XXS)Thinking can't be switched off | 509.2 s | 5 of 20 | 15 | 233.4 s |
128 GB · 1× DGX Spark (GB10)
128 GB class
| # | Setup | Finish time | Targets hit | Penalty loops | Thinking |
|---|---|---|---|---|---|
| 1 | ThinkingCap Qwen3.8-27B Kearuga1 DGX Spark 128 GB · TandemLLM 0.12.2 (0xBakeer, StairCut), community pack thinkingcap-27b-kearuga-exl3-4.00 (ThinkingCap Qwen3.8-27B Kearuga, EXL3 4.0 bpw), default reasoning effort xhighThinking off | 202.9 s | 13 of 20 | 7 | 32.1 s |
| 2 | ThinkingCap Qwen3.8-27B Kearuga1 DGX Spark 128 GB · TensorFold Python CUDA 0.6.6 (ashhart/TensorFold ed78d6fc + PR #260), 0xWhiteMage's ThinkingCap recipe @55efde56 (Qwen3.8-27B Kearuga, EXL3 4.0 bpw, DFlash2 drafter)Thinking off | 216.9 s | 12 of 20 | 8 | 34.1 s |
| 3 | Nemotron 3.5 Bolt 74B1 DGX Spark 128 GB · TensorFold, Mia's recipe (NVIDIA Nemotron 3.5 Bolt 74B IQ2, LiLiCorr draft)Thinking off | 266.1 s | 7 of 20 | 13 | 11.3 s |
| 4 | Gemma 4 31B1 DGX Spark 128 GB · TensorFold 1.2.5 (official release), Ash's recipe (Gemma 4 31B, MLX 4-bit, DFlash drafter)Thinking off | 274.3 s | 11 of 20 | 9 | 79.5 s |
| 5 | Qwen3.8-Flash-Next1 DGX Spark 128 GB · TensorFold, Mia's recipe (NVIDIA FP4 QAD, MTP)Thinking off | 286.3 s | 9 of 20 | 11 | 58.5 s |
| 6 | Ornith 1.5 122B1 DGX Spark 128 GB · TensorFold native (Zig @d1018b31 + Mia's patches 0001-0027), Mia's Ornith-1.5-122B recipe @0c8fd6a0 (EXL3 EAQ v1 4 bpw, DFlash2 drafts)Thinking off | 317.5 s | 5 of 20 | 15 | 35.7 s |
| 7 | Nemotron 3.5 Lightning1 DGX Spark 128 GB · TensorFold 1.2.5 (official release), Ash's recipe (NVIDIA Nemotron 3.5 Lightning 30B-A3B, MLX 4-bit, MTP)Thinking off | 342.0 s | 1 of 20 | 19 | 6.2 s |
| 8 | Ornith 1.5 122B1 DGX Spark 128 GB · TandemLLM 0.13.0 (0xBakeer, StairCut), Ornith-1.5-122B EXL3 4.0 bpw (0xBakeer pack, MTP drafts), default reasoning effort xhighThinking off | 377.7 s | 3 of 20 | 17 | 65.9 s |
256 GB · 2× DGX Spark (GB10)
256 GB class
| # | Setup | Finish time | Targets hit | Penalty loops | Thinking |
|---|---|---|---|---|---|
| 1 | DeepSeek V4.1 Flash2 DGX Sparks 256 GB · TensorFold, Mia's recipe (EXL3)Thinking off | 205.8 s | 12 of 20 | 8 | 23.0 s |
| 2 | DeepSeek V4.1 Flash2 DGX Sparks 256 GB · TensorFold native (Zig 0.17.0), Mia's recipe @0e7d421e (DeepSeek-V4.1-Flash-1017, EXL3 EAQ 2.9 bpw, DSpark drafts, tensor parallel 2)Thinking off | 206.2 s | 12 of 20 | 8 | 23.4 s |
| 3 | Qwen3.8-Flash-Next2 DGX Sparks 256 GB · TensorFold native (Zig 0.17.0, zig-flashnext db281878), Mia's Flash-Next Dual recipe @dc05c91 (NVIDIA Qwen3.8-Flash-Next NVFP4, MTP + copy drafts, tensor parallel 2)Thinking off | 283.2 s | 7 of 20 | 13 | 31.4 s |
| 4 | GLM-5.3-Flash2 DGX Sparks 256 GB · TensorFold 1.1.9, Mia's GLM-5.3-Flash 1104 recipe (d4226a1, EXL3 4 bpw, DFlash2), no prompt cacheThinking off | 300.8 s | 6 of 20 | 14 | 34.0 s |
| 5 | Qwen3.8-Flash-Next2 DGX Sparks 256 GB · TensorFold 1.2.5 (official release), Ash's recipe (NVIDIA Qwen3.8-Flash-Next NVFP4, tensor parallel 2)Thinking off | 345.8 s | 4 of 20 | 16 | 49.0 s |
| 6 | GLM-5.3-Flash2 DGX Sparks 256 GB · TensorFold native (Zig 0.17.0, engine f350cd3e + Mia's 112 patches), Mia's GLM-5.3-Flash recipe 1104 @d4226a1 (EXL3 4 bpw, DFlash2 drafts, tensor parallel 2)Thinking off | 346.3 s | 3 of 20 | 17 | 34.5 s |
| 7 | GLM-5.3-Flash2 DGX Sparks 256 GB · TensorFold Python engine v0.6.0 + 107 patches (Mia's pin image), Mia's GLM-5.3-Flash recipe @3fd60dfc (EXL3 4 bpw, DFlash2 drafts, tensor parallel 2)Thinking off | 352.9 s | 3 of 20 | 17 | 41.1 s |
| 8 | Ripple-65B2 DGX Sparks 256 GB · TensorFold, EXL3 5.0 bpw, DFlash2Thinking can't be switched off | 514.3 s | 5 of 20 | 15 | 238.5 s |
256 GB Mac · Mac Studio M5 Ultra 256 GB
256 GB Mac class
| # | Setup | Finish time | Targets hit | Penalty loops | Thinking |
|---|---|---|---|---|---|
| 1 | Gemma 4 31BMac Studio M5 Ultra 256 GB · TensorFold 1.2.5 (official release), Ash's recipe (Gemma 4 31B, MLX 4-bit, DFlash drafter)Thinking off | 210.5 s | 13 of 20 | 7 | 39.7 s |
| 2 | Ornith 1.5 122BMac Studio M5 Ultra 256 GB · TensorFold 1.2.5 (official release), Bob's recipe (Ornith 1.5 122B, 4-bit MLX)Thinking off | 290.4 s | 5 of 20 | 15 | 14.6 s |
| 3 | Qwen3.8-Flash-NextMac Studio M5 Ultra 256 GB · oMLX 0.7.0 (Jun Kim), MTP depth 3 (Qwen3.8-Flash-Next oQ4e requant by Jundot)Thinking off | 303.4 s | 5 of 20 | 15 | 15.6 s |
| 4 | Nemotron 3.5 LightningMac Studio M5 Ultra 256 GB · TensorFold 1.2.5 (official release), Ash's recipe (NVIDIA Nemotron 3.5 Lightning 30B-A3B, MLX 4-bit, MTP)Thinking off | 314.0 s | 3 of 20 | 17 | 8.2 s |
| 5 | DeepSeek V4.1 FlashMac Studio M5 Ultra 256 GB · TensorFold Metal (BobClawblaw's fork), BobClawblaw's recipe (DeepSeek-V4.1-Flash-1017 e2e3-p4-h8)Thinking off | 348.5 s | 5 of 20 | 15 | 63.7 s |
| 6 | LFM2.5-24B-A2BMac Studio M5 Ultra 256 GB · TensorFold 1.2.5 (official release), Ash's recipe (Liquid AI LFM2.5-24B-A2B, MLX 4-bit)Thinking off | 388.0 s | 0 of 20 | 20 | 37.2 s |
| 7 | Qwen3.8-Flash-NextMac Studio M5 Ultra 256 GB · inferd 0.4.1 (Incept5-AI, official release), author's agents-at-once recipe (Qwen3.8-Flash-Next oQ4e requant by Jundot, MTP)Thinking off | 395.3 s | 0 of 20 | 20 | 44.5 s |
Rank = order of the finish times of finished, verified races in the same memory class. A race that did not finish shows no number. Classes without a race are not listed. Click a race for the full result and every raw file.