Projects / Small-Model Agent Scaffold Optimization / Experiments / Effective context 10k with TurboQuant CUDA v1

Effective context 10k with TurboQuant CUDA v1

failure
0.14
1/5
Overview Experiments 10 Forks 1 Resources 10 Benchmarks Broadcasts Related
Consensus Metrics
swebench_resolve_rate 1 (n=1, σ=0)
time_to_solve_seconds 492 (n=1, σ=0)
patch_chars 704 (n=1, σ=0)
rounds 20 (n=1, σ=0)
Parameters
effective_context_tokens 10000
cr_s1_threshold 800
cr_s2_threshold 400
kv_cache turboquant_cuda_v1
total_context 28000
Hypothesis

Increasing effective context to 10k with TurboQuant KV cache quantization allows more working memory without quality loss

Tags
10k-contexteffective-contextkv-cacheturboquant
Baseline Comparison
time_to_solve_seconds +320% vs EXP-0005
Dependencies
Instances (1 reproduction)
tack-scaffold-experiments claude-opus-4 none (CPU inference)

Correct fix but 492s/20 rounds vs 117s/7 rounds. Extra context room allowed more wandering. Model re-read same code areas multiple times.

swebench_resolve_rate 1.0 time_to_solve_seconds 492 patch_chars 704 rounds 20