Local GLM vs the cloud

26 Aug 2026. Dual NVIDIA DGX Spark. Same 30-item hard suite, no tools. Cloud ox-alpha is full-precision GLM 5.3 Flash. Local is the NVFP4 quant plus a same-day SM120 NoPE overlay. Qwen3.8 Flash Next FP8 is the other Spark candidate. Flash stays parked until Léo says otherwise.

22P 5W 3Fox-alpha · cloud · 1272s

22P 4W 4FGLM NVFP4 · Sparks · 1775s

20P 6W 4FQwen FP8 · Sparks · 2073s

Same pass count as the cloud. One extra fail. Faster and cleaner than Qwen. NVFP4 beat the cloud on B-08 (which switch is really on) and H-05 (cite the live commit).

Thirty items, three seats

P passW weakF fail

	Q01	Q02	Q03	Q04	Q05	Q06	Q07	B-01	B-02	B-03	B-04	B-05a	B-05b	B-06	B-07	B-08	B-09	B-10	B-11	B-12	H-01	H-02	H-03	H-04	H-05	H-06	H-07	H-08	H-09	H-10
ox-alpha	P	P	P	P	P	P	P	W	F	P	P	P	P	P	W	F	P	P	P	P	P	W	W	F	W	P	P	P	P	P
GLM NVFP4	P	P	P	P	P	W	P	F	F	P	P	F	P	P	W	P	P	P	P	P	P	W	W	F	P	P	P	P	P	P
Qwen FP8	P	P	P	P	P	P	P	W	F	P	P	W	P	P	P	F	P	P	P	P	P	W	W	F	F	P	P	W	P	W

Fails

ox-alpha: B-02, B-08, H-04

GLM NVFP4: B-01, B-02, B-05a, H-04

Qwen FP8: B-02, B-08, H-04, H-05

Keiki coding (hard set)

Cloud ox-alpha swept this set (92/92 v1, 84/84 v2). Local GLM is the question. Auto-graded by Keiki's harness; P7 bug-list is still a manual field.

Files written: p07.py.

Finished: none. Errors/timeouts: P1, P2, P3, P4, P5, P6, P8, P9, P10.

# Scorecard — glm53fn

*Graded 2026-08-26 16:28 by Keiki's harness. Auto points below; MANUAL fields (quality 1/problem, P7 bug list 0-4) added by hand in the FINAL column.*

| Problem | Correctness | Perf | Manual | Final | Notes |
|---|---|---|---|---|---|
| p01 | 0.0 | — | q0-1 | _ | missing file |
| p02 | 0.0 | — | q0-1 | _ | missing file |
| p03 | 0.0 | — | q0-1 | _ | missing file |
| p04 | 0.0 | — | q0-1 | _ | missing file |
| p05 | 0.0 | — | q0-1 | _ | missing file |
| p06 | 0.0 | — | q0-1 | _ | missing file |
| p07 | 5.0 | — | 0-4 +q | _ |  |
| p08 | 0.0 | — | q0-1 | _ | missing file |
| p09 | 0.0 | — | q0-1 | _ | missing file |
| p10 | 0.0 | — | q0-1 | _ | missing file |

**Auto subtotal: 5.0 / (max auto: p07=5, p03/p04/p05/p08/p10=9, others=7 → 78)**

## Failure detail

Source: bakeoff-hard-suite scorecard 2026-08-26. Context on GLM: 131,072 tokens.