MiaAI-Lab/DeepSeek-v4.1-Flash-EXL3-2x-DGX-Sparks
DeepSeek v4.1 Flash EXL3 2.9 bpw for 2x DGX Sparks
Older snapshot: check the original source for current information.
From the publisher
README & documentation
Source previewRead the project’s overview, installation instructions and usage examples. The original README is the source of truth.
DeepSeek v4.1 Flash EXL3 2.9 bpw for 2x DGX Sparks by Mia'a AI Lab OpenAI-compatible vLLM serve of deepseek-ai/DeepSeek-V4.1-Flash as a local EXL3 2.9 bpw / mul1 checkpoint (model/, 39 shards, 196 GiB) on a 2× NVIDIA GB10 kit: tensor-parallel size 2 over CX7, native sm121a cubins, API on :8888. Served model id: DeepSeek-v4.1-Flash-EXL3. Speculation is DSpark, and the draft experts live in the checkpoint (mtp., dsparkblocksize=5, 128 draft experts / top-3, target layers 37–39) — there is no separate drafter to download. The image overlays EXL3 onto vllm/vllm-openai:deepseekv41-flash-0909 (linux/arm64, vLLM 0.1.dev20904+g179dd0fa9), the only base carrying the DeepseekV41 architecture. Host RAM is GPU memory on a Spark, and every cudaMalloc is committed the moment it is made. Per node, from scripts/weightbudget.py --tp 2: Shipped defaults: MAXMODELLEN=600000, MAXNUMSEQS=2, MAXNUMBATCHEDTOKENS=1536, DSV41IOTHREADS=96, and a 2.5 GiB KV pool. The 1536/96 pair is the measured agent-workload setting; the previous 1024/32 pair remains an easy…
Read the full README ↗ · Preview checked 2026-10-31T16:44:22.930Z
Inside the original README — Document outline
- What this checkpoint is
- Speculation
- Memory
- Memory guard — off by default, and not needed in practice
- Quick start
- Optional: Keys abliterated overlay (third party)
- Optional: cooperative MoE (decode, two Sparks)
- Images
- Layout
- Engram
- Agent prefix caching
- CX7
Headings are captured from the source. Links open the publisher’s document, not a locally hosted copy.
Links from the README
References supplied by the publisher, not independently verified endorsements. Check the destination before downloading files or entering credentials.
Documentation belongs to its respective authors. Reported project/model license: AGPL-3.0. A listing is not a grant of reuse or training rights. Confirm the document’s own terms at the source.
What this repository does
DeepSeek v4.1 Flash EXL3 2.9 bpw for 2x DGX Sparks
Repository facts
- Owner
- MiaAI-Lab
- Primary language
- Python
- Stars
- 221
- Forks
- 30
- Open issues + pull requests
- 14
- License
- AGPL-3.0
- Archived
- No
- Default branch
- main
- Created
- 2026-09-13T09:10:00.000Z
- Last push
- 2026-09-19T13:38:32.000Z
Topics and intended use
No topics were included in the latest source metadata.
Review the README for scope, installation, examples and limitations. We do not run repository code or certify it.
Evaluate before installing
Review licensing and dependencies, inspect recent commits and unresolved issues, and test in an isolated environment before production use. Stars and forks alone cannot answer these questions.
Source and freshness
Source: GitHub. Metadata observed 2026-09-20T06:40:45.138Z. Daily imports are snapshots, not real-time monitoring.
Popularity and source listings do not establish security, suitability, licensing rights or benchmark performance.
Related repositories
EbookFoundation/free-programming-books
donnemartin/system-design-primer