DeepSeek V4 Flash — REAM96 (111B) · DS4 Q2
A 2-bit build of DeepSeek-V4-Flash-0731-REAM96-111B — DeepSeek-V4-Flash with 96 of the original 256 experts per layer, sized for 48 GiB configurations. Quantized with the standard DS4 recipe (2-bit experts, 8-bit attention) and a fresh importance matrix.
[!IMPORTANT] This is a DS4-specific GGUF. Run it with the DS4 fork — the 96-expert topology needs its variable expert count support. Generic llama.cpp will not load this file.
[!WARNING] Live smoke testing passed 5/10 scenarios on the first run. Independent reruns show the failures (Russian wordplay, multi-turn, English → Russian code-switching, Tool calling (DSML), Long-dialog focus (drift check), Tool call → code chain) are intermittent, not absolute — see the Stability column below for per-scenario pass rates. Multilingual chat, reasoning and long dialogs are consistently healthy. The full-precision native checkpoint may behave better — 2-bit quantization hits agentic behavior hardest.
Files
| File | Size | What it is |
|---|---|---|
ream96.gguf | 36.4 GB | The model |
dspark.gguf | 5.7 GB | Optional speculative decoding (DSpark) |
imatrix.dat | 0.2 GB | Importance matrix used for this quant |
SMOKE_REPORT.json | — | Raw smoke-test evidence |
Run
Plain:
./ds4 -m ream96.gguf -c 8192
With DSpark speculative decoding (faster generation, more memory):
./ds4 -m ream96.gguf --mtp dspark.gguf --dspark -c 8192
DSpark fits alongside on 64 GiB hosts, but in our Apple Silicon runs it gave no measurable speedup (12.9 vs 13.0 tok/s) — treat it as experimental.
How it was made
One pruning step, straight from the original — no cascading. Expert importance was
measured by running deepseek-ai/DeepSeek-V4-Flash-0731
over a ~5-million-token calibration mix (multi-turn dialogs, thinking and direct modes,
rendered with the model's own chat encoder). The strongest experts of every domain were
protected from pruning, the survivors were carried over byte-identical, and the
router was re-balanced to keep the original selection behavior.
| Calibration domain | Share |
|---|---|
| Code | 35% |
| Agentic / tool use | 19% |
| Multilingual chat | 16% |
| Math | 8% |
| General chat | 6% |
| Roleplay | 6% |
| Russian | 5% |
| Long docs | 4% |
This line replaces the earlier cascaded REAM builds (now archived under -exp names),
which degraded badly in multi-turn use.
Smoke results
Every scenario is a live multi-turn conversation run end-to-end on the DS4 runtime (raw evidence ships in SMOKE_REPORT.json).
| Scenario | First run | Stability (reruns) |
|---|---|---|
| Russian wordplay, multi-turn | ❌ | 2/10 |
| English → Russian code-switching | ❌ | 0/10 |
| Code Q&A over a 4k-token file | ✅ | — |
| Tool calling (DSML) | ❌ | 0/10 |
| Russian multi-turn reasoning | ✅ | — |
| Spanish creative writing | ✅ | — |
| Code refactoring | ✅ | — |
| Chinese summarization | ✅ | — |
| Long-dialog focus (drift check) | ❌ | 2/10 |
| Tool call → code chain | ❌ | 0/10 |
Stability = pass rate over independent reruns of the scenarios that failed the first run; passing scenarios were not re-run.
Limitations
- Needs the DS4 fork; not a generic llama.cpp file.
- 2-bit quantization is aggressive: expect the native checkpoint to be smarter than this build, especially on agentic tool use.
- Memory use grows with context length and DSpark; the sizes above are the files alone.