bigbang-v1-trial

endless-frontier/BigBang-v1 — general-purpose post-train of Qwen/Qwen3.6-35B-A3B

via an adversarial, self-evolving synthetic-data framework over VERIFIABLE

frontier tasks (formal methods, computation, simulation, domain tools).

apache-2.0, 221 likes, bartowski GGUF 562k downloads. Trialled alongside

[[xyz-aquila-mini-trial]] as a candidate for tommy's slot

([[thomson-1.0-small-trial]]). Served as bigbang on port 9176.

VERDICT: REJECTED. Dominated by aquila on nearly every axis, and by tommy on

all the ones that matter. Weights are pure disk cost — 23.74 GiB Q5_K_M + 24.03

GiB Q5_K_L on /mnt/models.

The context penalty comes first

Its Q5_K_M is 23.74 GiB — only 450 MiB more than aquila's 23.30 — and that

450 MiB is the difference between holding 262144 and not. At 262144 it LOADS,

reports /health ok, then hard-crashes on the first image (GGML_ASSERT in

ggml_backend_tensor_alloc). It must run at 196608 to afford the lazy ViT

buffer. So it pays 65k of context for its extra bits before scoring a point.

Full numbers in [[qwen35moe-35b-vram-ceiling]].

Grill (bench/grill-bigbang-20260830-1103, same harness/day as aquila)

| suite | bigbang | aquila | tommy |

|---|---|---|---|

| coding r1-r4 | 20/23 (worst) | 22/23 | 21/23 |

| R5 complete | 4/6 | 3/6 | 6/6 |

| R5 looped | 1/6 (multifile_refactor) | 1/6 | 0/6 |

| R5 redundant | 17.4% (worst) | 11.1% | 0.0% |

| klayout unaided | 0/8 | 1/8 | 1/8 |

| klayout + ref | 7/8 | 6/8 | 7/8 |

| klvision | 5/6 | 5/6 | 5/6 |

| closedloop unaided | 0/4 | 0/4 | 4/4 |

| closedloop + ref | 4/4 | 4/4 | 4/4 |

| office unaided | 3/9 | 8/9 | 9/9 |

| office + ref | 3/9 | 6/9 | 9/9 |

| realcase | 7/7 PASS | 7/7 PASS | 0/7 (GENUINE fail, re-tested 08-30) |

| envelope | CLEAN, 9.0% abnormal | CLEAN, 4.8% | CLEAN |

Combined on tommy's denominator: bigbang 24/29, aquila 25/29, tommy 27/29.

(The three R5 rows above are the single battery run; the x3 re-run below supersedes them.)

R5 RE-RUN x3 (bench/grill-r5x3-20260830-1429) — LOOPS ON TWO TASKS

n=4 (battery + 3 reps):

| run | complete | looped | redundant | which tasks |

|---|---|---|---|---|

| battery | 4/6 | 1/6 | 17.4% | multifile_refactor |

| rep1 | 4/6 | 1/6 | 10.5% | interrupt_replan |

| rep2 | 3/6 | 2/6 | 19.6% | multifile_refactor + interrupt_replan |

| rep3 | 3/6 | 2/6 | 20.0% | multifile_refactor + interrupt_replan |

**bigbang loops on TWO distinct tasks — interrupt_replan 3/4 runs AND

multifile_refactor 3/4 runs** — where [[xyz-aquila-mini-trial]] loops on exactly

one (interrupt_replan, 4/4). Mean completion 3.5/6 vs aquila's 3.75/6, and

redundancy is the worst of the three models in every single run (10.5-20.0%).

So on the [[grill-round5-agentic-loops]] dealbreaker axis bigbang is the worst of

the three at n=4, confirming the battery's single sample rather than rescuing it.

THE DISQUALIFIER: office 3/9 in BOTH arms — a capability ceiling

This is the finding worth keeping. line_months:struct and

formula_totals:struct fail identically with and without the API reference

(771/718 tok unaided, 425/510 tok with ref — it fails FASTER with the docs, not

better), each cascading to 2 BLOCKED stages.

That is a third, worse failure mode than the two already catalogued:

• [[tommy2-iq2m-grill]] / [[gemma12-grill]]: recall BROKEN, reference FIXES it

(0/9 -> 9/9). Fixable by loading a skill.

• [[xyz-aquila-mini-trial]]: recall INTACT on office (8/9 unaided), missing only

on niche KLayout. Fixable, domain-scoped.

• bigbang: the reference makes NO difference. It is not missing the API, it

cannot use it. Nothing you put in context fixes that.

Only bar_regions passes, so it clears just 1 of the 3 office scenarios.

(Note [[office-suite-vision-floor]]: 3 of 9 stages need vision, so a text-only

model caps at 6/9 — bigbang has working vision and still lands at 3/9.)

Other negatives

• klayout unaided 0/8 — the only model of the three to score zero.
• 9.0% abnormal finish_reason (12 of 133 turns), double aquila's 4.8%.
• R5 redundancy worst of the three in EVERY run (10.5-20.0%). Its battery

4/6 looked like it edged aquila's 3/6, but at n=4 the means invert

(3.5 vs 3.75) and bigbang loops on two tasks to aquila's one.

What it got right (not enough)

realcase 7/7 genuine, closedloop+ref 4/4, klayout+ref 7/8 (ties tommy), klvision

5/6, envelope CLEAN with zero control-token leaks. It is not a broken model — it

is a competent one with a hard ceiling on structured-output/office work and no

advantage anywhere that would justify the slot.

Gotchas

• GGUF ships MTP / nextn tensors (blk.40.nextn.*) that llama.cpp logs as

unused tensor -- ignoring. Harmless, but the published MTP speedup is NOT

active on llama.cpp — do not expect it (cf. [[qwen38-atomic-install]], where MTP

was undocumented and had to be probed).

• Triage check 8 ([[model-triage-checklist]]): derivative of a served base, must

justify on the grill. It did not.