Releases·Tool Release
DeepSeek V4.1F Q4 on M3 Ultra with native DSpark MTP (40tps / 800tps)
A developer forked antirez/ds4 to optimize DeepSeek V4.1 Flash Q4 inference on Apple's M3 Ultra, roughly doubling decode speed and adding the first DSpark multi-token-prediction implementation on Metal. The fork preserves byte-identical output under greedy decode and is published with verification manifests on GitHub.
58 engagement1 sourceTue, Sep 15, 2026, 11:51 AM
AI Overview
- A fork of antirez/ds4 runs DeepSeek V4.1 Flash Q4 on a 512 GB M3 Ultra, doubling decode from 16.6 to 31.3 t/s at 8k context.
- Long-context decode hits 28.3 t/s at 300k tokens, 90% of the 8k rate, by pruning the attention block-selection chain.
- Native DSpark multi-token speculation on Metal lifts decode to 40.5-41.3 t/s, with byte-identical greedy output to upstream.
- Disk KV cache restores a full prefix in 0.23 s vs 31 s cold; prefill gains only 10% because GEMM already runs near its ~23-24 TFLOPS ceiling.
- The M3 Ultra-only branch needs ~300 GB RAM, redistributes no weights, and reports a 91-minute agent turn with 56 error-free tool calls.
Related
Social·Community
Running a 341 GB model on a 128 GB Mac: two small macOS fixes made the first token 8x faster
A developer detailed two macOS fixes in sf-ds4-1flash, a fork of antirez's ds4 limited to DeepSeek V4.1 Flash on Metal, which streams a 341 GB Q2 GGUF from SSD on a 128 GB M5 Max. Fixing APFS page-boundary caching behavior and switching to a GPU mailbox poll made the first token after an 8K prompt drop from 1.23 s to 0.15 s. The techniques apply to anyone streaming model weights from disk on macOS.
A developer detailed two macOS fixes in sf-ds4-1flash, a fork of antirez's ds4 limited to DeepSeek V4.1 Flash on Metal, which streams a 341 GB Q2 GGUF from SSD on a 128 GB M5 Max. Fixing APFS page-boundary caching behavior and switching to a GPU mailbox poll made the first token after an 8K prompt drop from 1.23 s to 0.15 s. The techniques apply to anyone streaming model weights from disk on macOS.
Updated Sun, Oct 11, 2026, 08:17 PM — Two follow-up fixes landed after the initial fork post: APFS boundary-aligned reads and GPU mailbox polling, lifting pre
8 engagement1 sourceredditSun, Oct 11, 2026, 08:17 PM
Social·Community
Developer speeds DeepSeek V4 Flash 12x on M3 Ultra with sparse-attention kernel PRs
A developer serving DeepSeek V4 Flash on a Mac Studio M3 Ultra (512GB) via the DwarfStar (ds4) project cut chat-turn latency from 6–20 seconds to 1.6 seconds. Three stacked pull requests optimize the sparse-attention lightning indexer, achieving a 12x speedup with bit-exact results and a 21% cold-prefill improvement at 64k context.
A developer serving DeepSeek V4 Flash on a Mac Studio M3 Ultra (512GB) via the DwarfStar (ds4) project cut chat-turn latency from 6–20 seconds to 1.6 seconds. Three stacked pull requests optimize the sparse-attention lightning indexer, achieving a 12x speedup with bit-exact results and a 21% cold-prefill improvement at 64k context.
34 engagement2 sourcesredditWed, Aug 19, 2026, 12:48 AM
Releases·Tool Release
Redis creator launches ds4, a tool for running LLMs locally
From the creator of Redis: ds4 runs LLMs locally
From the creator of Redis: ds4 runs LLMs locally
402 engagement1 sourcehackernewsFri, Oct 2, 2026, 06:01 PM
Products·building
ds4: high-throughput local inference for GLM-5.3-Flash on M3 Ultra (60-550 tps, Q4)
ds4 is a decoding setup for running GLM-5.3-Flash locally on Apple M3 Ultra hardware, pushing Q4 throughput from ~59% to a targeted 80% of measured memory bandwidth. It enables the model to feel responsive inside a Claude Code harness at ~200k context depth with many tool calls. Aimed at local LLM enthusiasts who need materially faster output from GLM-5.3-Flash.
ds4 is a decoding setup for running GLM-5.3-Flash locally on Apple M3 Ultra hardware, pushing Q4 throughput from ~59% to a targeted 80% of measured memory bandwidth. It enables the model to feel responsive inside a Claude Code harness at ~200k context depth with many tool calls. Aimed at local LLM enthusiasts who need materially faster output from GLM-5.3-Flash.
144 engagement1 sourceredditWed, Sep 9, 2026, 12:51 PM
Social·Community
DeepSeek V4.1 Flash on a single DGX Spark: 113.6 GB VQ base + 40 MB domain sidecars, 74–82% top-1 agreement vs original
A developer released details of YoungAi, a native C/CUDA inference engine that runs DeepSeek V4.1 Flash on a single NVIDIA DGX Spark (GB10, 128 GB unified memory) despite the original weights being ~510 GB. It quantizes experts into a 113.6 GB zero-corpus VQ base and adds ~40 MB per-domain 'sidecar' files that restore domain behavior. Sidecars raise teacher-forced top-1 agreement with the original model by +2.8 to +3.7 points across five domains, with speculative decoding reaching 43 tok/s on a real agent request.
A developer released details of YoungAi, a native C/CUDA inference engine that runs DeepSeek V4.1 Flash on a single NVIDIA DGX Spark (GB10, 128 GB unified memory) despite the original weights being ~510 GB. It quantizes experts into a 113.6 GB zero-corpus VQ base and adds ~40 MB per-domain 'sidecar' files that restore domain behavior. Sidecars raise teacher-forced top-1 agreement with the original model by +2.8 to +3.7 points across five domains, with speculative decoding reaching 43 tok/s on a real agent request.
2 engagement1 sourceredditTue, Oct 6, 2026, 11:30 AM