82 Billion Tokens. One GPU. 2.7× Faster.

Verbatim Reproduction (MRCR v2)

MRCR tests whether a model can find and reproduce exact text from long conversations. The same request is repeated 2-8 times — the model must identify and reproduce the Nth instance verbatim. This is OpenAI's benchmark for long-context quality.

A 35B open-weight model on a single RTX 3090 outperforms GPT-5.5, Gemini 3.1 Pro, and Claude Opus 4.7 at every context length tested. 99.5% accuracy up to 512K. At 512K-1M, AlphaChat scores 90.7% — beating GPT-5.5 (74%) by +17pp, Opus 4.7 (32.2%) by +59pp, and Gemini 3.1 (26.3%) by +64pp.

MRCR uses SequenceMatcher scoring (0-1 scale). Frontier model scores from published benchmarks (benchlm.ai, llm-stats.com, Anthropic system card). “—” = not publicly benchmarked at that context range. AlphaChat scores measured on identical dataset, same scoring method, using sequential context mode (--context sequential).

Needle in a Haystack: Standard Retrieval vs AlphaChat

10 tests per category, 187 sub-scores per scale. Standard keyword retrieval (BM25) degrades at scale. AlphaChat doesn't.

AlphaChat peaks at 98.9% at 0.4B tokens, +68pp over standard retrieval. Maintains 92-99% from 4 million to 82 billion tokens — scale-invariant. Standard keyword retrieval collapses from 44% to 28%.

Code Dependency Understanding (DependEval DR)

DependEval tests whether a model can trace cross-file dependencies in real codebases — which files import from which, what classes depend on what. This is the core skill needed for code navigation, refactoring, and bug tracing. Benchmark: DependEval (ACL 2025), Dependency Recognition task. 2,683 repositories across 8 programming languages.

A 35B model on one GPU outperforms every tested model including DeepSeek-V3 (671B, 19× larger) and all 2026 frontier APIs. AlphaLlama's extended code mode understands cross-file dependencies that raw models miss.

Retrieval at Repository Scale (DependEval Haystack)

Real codebases are large. We tested what happens when you scale DependEval from its original 3K tokens to real repository sizes (10K → 1M tokens).

Raw Model Performance — Every Model Collapses

With AlphaLlama Extended Code Mode — Scale-Invariant

Extended code mode provides scale-invariant retrieval. While raw models degrade and collapse, AlphaLlama maintains 70-90% accuracy regardless of codebase size.

Key discovery: retrieval ability ≠ comprehension ability. GPT-5.5 is the best at finding code (100% to 200K) but the worst at understanding dependencies (30% DependEval DR). AlphaLlama is the best at understanding (82.5% DR). Extended code mode combines both strengths.

DependEval-Haystack is our benchmark extension. 10 samples per scale, reproducible.

Inference Speed: AlphaLlama vs Ollama

Same model, same GPU, same quantization. AlphaLlama is 2.1-4.6× faster on oversize MoE models thanks to smart offloading.

Test: Qwen3.6-35B (Q4_K_M, 22 GB) on RTX 4080 (16 GB VRAM). Model exceeds GPU memory — overflow handling determines speed.

Ollama version: 0.20 (latest, June 2026). When the model fits entirely in VRAM, both engines perform equally. Speed advantage applies to MoE models that exceed VRAM.

AlphaChat vs SubQuadratic SubQ

SubQuadratic raised $29M and launched SubQ with a 12M token context window.

Speed & Storage

VRAM and RAM stay constant regardless of context size. Scaling is disk-only.

~1s is per-query latency against the pre-indexed corpus, not processing 82B tokens live.

Benchmark Methodology

All benchmarks use synthetic needle facts — unique strings that do NOT exist in the model's training data. The model must find the fact in the corpus, not recall it from memory. 6 categories tested, 10 tests per category, 187 sub-scores per scale, 7 scales from 4M to 82B tokens.

82B Token Test Corpus

82 billion tokens across 14 datasets spanning encyclopedic knowledge, code, mathematics, and multilingual text.

Why Unlimited Context Matters