◂ BACK TO LEADERBOARD
▚ MODEL AUDIT // GGUF

// Muse Glimmer 28B

FLAWS — Minor Logic Flaws 76/100 tests: CRASH
⚠ SPEED CAVEAT: Slow deep-thinker: ~17-19 t/s, 5-9 min/prompt, ~5-9k tokens incl. reasoning. Fast for quality, not throughput.
tok/sec19.38
tokens9095
TTFT3.55s

▮ PILLAR BREAKDOWN

Complexity (O(1))
18/20
Concurrency / Races
15/20
Tx Isolation
12/20
Memory & Edges
16/20
Test Integrity
15/20

✓ WENT RIGHT

  • Complexity (O(1)) (18/20)
  • Memory & Edges (16/20)

✗ WENT WRONG

  • Tx Isolation (12/20)

▮ CRITICAL BUGS

▮ RECOMMENDED USE

The strongest LFU result from a non-coder model in this benchmark — nails true O(1) freq-bucket eviction with correct min_freq tracking and __slots__ where both specialist coder models (KAT, Qwen3-Coder) failed catastrophically. Let down by one signature bug: rollback() is sync but awaited.

▮ REFACTORED PATCH

# Fix 1 (fatal): rollback must be async to match 'await tx2.rollback()'
async def rollback(self):
    self.pending_puts.clear()
    self.pending_deletes.clear()

# Fix 2: bound _update_min_freq against a drained cache
def _update_min_freq(self):
    if not self.freq_map:
        self.min_freq = 0; return
    while self.min_freq not in self.freq_map:
        self.min_freq += 1

# Fix 3: lifecycle guard on Transaction
def __init__(self, cache):
    self.cache = cache; self.pending_puts = {}; self.pending_deletes = set(); self._closed = False
def _check_open(self):
    if self._closed: raise RuntimeError("Transaction already committed/rolled back")
async def commit(self):
    self._check_open(); # ... existing body ...; self._closed = True
async def rollback(self):
    self._check_open(); self.pending_puts.clear(); self.pending_deletes.clear(); self._closed = True