Recipe
deepseek-v4-flash-iq3-xxs-rtx-pro-6000-blackwell-96gb-llama-cpp-tp1
deepseek-v4-flash-iq3-xxs-rtx-pro-6000-blackwell-96gb-llama-cpp-tp1Observed LocalMaxxing leaderboard run. Evidence for compatibility, not an executable launch contract.
Record
- Status
- candidate
- Source
- localmaxxing
- Engine
- llama.cpp
- Engine version
- dsv4 fork + in-graph K=1 MTP + prefill kernel patches (batched indexer top-k, compact MoE work-list)
- Accelerators
- 1
- Tensor parallel
- 1
- Context tokens
- 262,144
- Max concurrency
- 1
- chat
- unknown
- reasoning
- unknown
- tools
- unknown
- vision
- unknown
Hugging Face model card
Identity
https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash- Repository
- deepseek-ai/DeepSeek-V4-Flash
- Status
- known
- Link type
- Exact Hub repository
Public Hugging Face repository confirmed by the Hub API.
Candidate: useful compatibility or speed evidence. The registry does not offer Run until promotion requirements are met.
Observed configuration
llama.cpp
Evidence only · candidate · reference
Candidate evidence — not a Run contract
Observed source tokens
Mechanical split of the source command. Unverified against the engine CLI. The registry does not offer Run for this recipe.
llama-server-mDeepSeek-V4-Flash-custom-iq3down.gguf-c262144-ngl99-np1-b2048-ub512--spec-typedsv4-mtp--cache-reuse256--no-mmap--flash-attnon
| Flag | Value |
|---|---|
-m | DeepSeek-V4-Flash-custom-iq3down.gguf |
-c | 262144 |
-ngl | 99 |
-np | 1 |
-b | 2048 |
-ub | 512 |
--spec-type | dsv4-mtp |
--cache-reuse | 256 |
--flash-attn | on |
Environment
| Variable | Value |
|---|---|
DSV4_MTP_GGUF | DeepSeek-V4-Flash-MTP-Q4K-Q8_0-F32.gguf |
DSV4_MTP_SPEC | 1 |
Measured speed
| Concurrency | Context | Prefill | Decode | TTFT ms | Status | Sweep |
|---|---|---|---|---|---|---|
| 1 | 262,144 | 1,688 | 80.8 | 9,240 | observed | deepseek-v4-flash-iq3-xxs-rtx-pro-6000-blackwell-96gb-llama-cpp-tp1-sweep |
Remaining fields
Identity, launch, related records, and measured speed are shown above. This is the rest of the normalized record.
- hardware count
- 1
- hardware id
- rtx-pro-6000-blackwell-96gb
- id
- deepseek-v4-flash-iq3-xxs-rtx-pro-6000-blackwell-96gb-llama-cpp-tp1
- model instance id
- deepseek-ai-deepseek-v4-flash--iq3-xxs
- recipe source
- localmaxxing
- schema version
- local-ai-registry/v1
- speed sweep ids
- deepseek-v4-flash-iq3-xxs-rtx-pro-6000-blackwell-96gb-llama-cpp-tp1-sweep
- status
- candidate
capabilities
engine
- name
- llama.cpp
- version
- dsv4 fork + in-graph K=1 MTP + prefill kernel patches (batched indexer top-k, compact MoE work-list)
serving
- max concurrency
- 1
- max context tokens
- 262,144
- tensor parallel
- 1
Provenance & metadata (3)
facts
- capabilities.chat · provenance · captured at
- 2026-08-30T09:10:02Z
capabilities.chat · reason capability-not-verified
capabilities.chat · state unknown
- capabilities.reasoning · provenance · captured at
- 2026-08-30T09:10:02Z
capabilities.reasoning · reason capability-not-verified
capabilities.reasoning · state unknown
- capabilities.tools · provenance · captured at
- 2026-08-30T09:10:02Z
capabilities.tools · reason capability-not-verified
capabilities.tools · state unknown
- capabilities.vision · provenance · captured at
- 2026-08-30T09:10:02Z
capabilities.vision · reason capability-not-verified
capabilities.vision · state unknown
- engine.graph mode · provenance · captured at
- 2026-08-30T09:10:02Z
engine.graph mode · reason runtime-detail-not-published
engine.graph mode · state unknown
- serving.kv cache tokens · provenance · captured at
- 2026-08-30T09:10:02Z
serving.kv cache tokens · reason kv-cache-capacity-not-published
serving.kv cache tokens · state unknown
- serving.max concurrency · provenance · captured at
- 2026-08-31T23:03:15Z
serving.max concurrency · reason server-capacity-derived-from-source-evidence
serving.max concurrency · state known
metadata
- localmaxxing · backend
- cuda
- localmaxxing · batch size
- 1
- localmaxxing · hardware label
- RTX PRO 6000 Blackwell
- localmaxxing · notes
- Full 284B-A13B MoE resident on ONE 96GB card with in-graph K=1 MTP speculative decoding: the DS4 graph drafts its own next token each decode (antirez's 3.8GB MTP module side-loaded) and the server verifies it in a constant nt=2 batch. Draft acceptance 75-83%, ~1.45x over the same quant without MTP (42-47 t/s). Decode is steady-state from a 10-hour soak: 80.8 t/s held across a continuous 47K-token generation at temp 0. Prefill re-measured after two custom CUDA prefill patches in the fork (batched lightning-indexer top-k, fixing a >1M-kernel launch storm; compact MoE stream-k work-list, fixing last-wave quantization): 15.6K-token prompt at 1688 t/s (was 486 pre-patch, ~3.5x). Custom mixed quant (2.58bpw): antirez q2 imatrix recipe with ffn_down_exps upgraded Q2_K -> IQ3_XXS (wikitext PPL 4.505 vs 4.717 stock). Fork lineage: fairydreaming dsv4 graph/kernels + K=1 MTP ported from alesha-pro ds4-longctx. HumanEval+ pass@1 93.1% (official localmaxxing eval shard run, this quant, temp 0). Aider polyglot (Python, diff): pass@2 64.7%. Fork: github.com/sigkaiser/llama.cpp (mtp-k1 branch). Quant: hf.co/starvingcheetah/DeepSeek-V4-Flash-custom-iq3down-GGUF.
- localmaxxing · observed command
- DSV4_MTP_GGUF=DeepSeek-V4-Flash-MTP-Q4K-Q8_0-F32.gguf DSV4_MTP_SPEC=1 llama-server -m DeepSeek-V4-Flash-custom-iq3down.gguf -c 262144 -ngl 99 -np 1 -b 2048 -ub 512 --spec-type dsv4-mtp --cache-reuse 256 --no-mmap --flash-attn on
- localmaxxing · run id
- cmrch4u4u00hro201qqq1pwiz
- localmaxxing · tokenized · arguments
- llama-server, -m, DeepSeek-V4-Flash-custom-iq3down.gguf, -c, 262144, -ngl, 99, -np, 1, -b, 2048, -ub, 512, --spec-type, dsv4-mtp, --cache-reuse, 256, --no-mmap, --flash-attn, on
- localmaxxing · tokenized · environment · DSV4 MTP GGUF
- DeepSeek-V4-Flash-MTP-Q4K-Q8_0-F32.gguf
- localmaxxing · tokenized · environment · DSV4 MTP SPEC
- 1
- localmaxxing · tokenized · fidelity
- faithful
provenance
- captured at
- 2026-08-30T09:10:02Z
sources
| captured at | kind | url |
|---|---|---|
| 2026-08-30T09:10:02Z | normalized-recipe | www.localmaxxing.com/en/runs/cmrch4u4u00hro201qqq1pwiz ↗ |