| # Quantization Study — Qwen3.8-Flash-Next (brief §13) |
|
|
| Date 2026-11-03. Both builds served by the same llama.cpp (29c720e), same 288K strict-256K config (YaRN ×2, f16 KV, routed experts + n-gram table in RAM), same harness and seeds. All values MEASURED. |
|
|
| ## Builds |
|
|
| | | Unsloth UD-Q4_K_XL | ISTA-DASLab GSQ-RCO IQ3_S | |
| |---|---|---| |
| | Size on disk | 111.3 GB (4 shards) | 83.6 GB (2 shards) | |
| | Transformer weights (must be resident) | ~82 GB | 54.8 GB | |
| | n-gram table (memory-mapped) | 28.8 GB IQ4_NL | 28.8 GB IQ4_NL (same) | |
| | Method | dynamic per-tensor 4/5/8-bit (Q4_K / Q5_1 experts) | learned scalar quantization + Riemannian budget allocation, ~3.4 bpw on experts | |
| |
| The ISTA unpruned GSQ-RCO release was used (all 512 experts); the expert-pruned "Coder" variant was not tested. |
| |
| ## Quality (no perplexity — task results only, as the brief requires) |
| |
| | Test | Q4 | IQ3_S | |
| |---|---|---| |
| | Shared 262K prompts v2 (same text Opus/MiniMax got), 13 checks each | 13/13, 13/13, 13/13 | 13/13, 13/13, 13/13 | |
| | Cross-file aggregation F1 on those prompts | 1.0 / 1.0 / 1.0 | 1.0 / 1.0 / 1.0 | |
| | AIME 2026, 15 odd-numbered problems | 14/15 | 13/15 | |
| | AIME per-problem differences | — | #11: IQ3_S hit the 32K token cap; Q4 solved it. All other outcomes identical (both miss the same one). | |
| |
| Interpretation: long-context retrieval and cross-context reasoning are unaffected at 262K. Hard math shows a possible small loss (one problem; within noise for n=15, but in the expected direction). |
| |
| ## Speed and memory |
| |
| | | Q4 | IQ3_S | Change | |
| |---|---|---|---| |
| | Short prompt (~4K): prefill / decode | 866 / 33.7 tok/s | 1,042 / 39.0 tok/s | +20% / +16% | |
| | 262K prompt: prefill | ~762 tok/s (5.7–5.9 min) | ~916 tok/s (4.8 min) | +20% | |
| | 262K: decode | 29.7 tok/s | 33.8 tok/s | +14% | |
| | Peak VRAM (288K window) | 19.2–19.4 GB | 18.2 GB | −1 GB | |
| | Resident RAM for weights | ~82 GB | ~55 GB | −27 GB | |
|
|
| ## Recommendation |
|
|
| - **Default: Q4 (UD-Q4_K_XL).** It is the build that passed the full 10-run 256K qualification and scored highest on AIME. |
| - **Speed profile: IQ3_S.** Use when throughput or RAM matters more than the last bit of hard-reasoning quality (agentic loops, or running other RAM-heavy services alongside). Label **CONDITIONAL**: passed 3/3 strict-256K runs but has not run the 10-run stability suite. |
| - The 27 GB of RAM IQ3_S frees is enough to keep Ollama models or another service resident without memory pressure. |
| |
| ## Not tested (and why) |
| |
| - Q5/Q6/Q8 references: Q5_K_XL needs ~129 GB resident for the transformer weights; the box has 125 GB. Q4 is the highest precision that fits, so it served as the reference. |
| - IQ2-class builds (ISTA IQ2_XS 68 GB, Unsloth Q2_K_XL 79 GB): not tested; IQ3_S already showed a first sign of loss on hard math, so lower bits are unlikely to be worth it for a quality-first brief. |
| |
