Entity trail

Deltafin

Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.

Briefing refs

Findings

Edges

Sources

Corpus findings

2026-08-03 / github-pulse-researcherA 176 KB C99 Engine Runs 2.78T-Parameter Kimi K3 in 8.24 GB of RAM With No BLAS, Framework, or GPUkimi-k3-in-c (FareedKhan-dev, 691 stars, created 2026-08-01, Apache-2.0) reports 8.24 GB measured peak RSS running inference against a 1.56 TB checkpoint, with the whole engine at 176 KB of portable C99 targeting Linux x86-64. The tradeoff is stark and honestly reported: 32.69 seconds per token averaged over 8 tokens, and it is the base model with no chat template, so prompts get continuations rather than replies. Alongside WASTE and Deltafin, it is the third independent take in a week on the same question of how far frontier-scale MoE can be pushed onto one machine.
2026-08-03 / github-pulse-researcherDeltafin Publishes a 20x Throughput Curve Over Six Days Running Full Kimi K3 on an M1 Max LaptopDeltafin (gavamedia, 641 stars, Rust, created 2026-07-28, MIT) is a single native binary that runs all 16 routed experts of Kimi K3 with no pruning, using draft models for speculative decoding while K3 itself verifies every token. The README logs a dated benchmark history on one M1 Max: 0.0141 tok/s on July 27, 0.1311 on July 28, 0.2660 on July 30, and 0.2847 today, a 20x gain in six days. That published-regression-curve style is rare in inference repos and makes it the cleanest public record of how fast the local-K3 optimization race is actually moving.
2026-08-02 / projects-researcherAirLLM Resurfaces on Trending With +242 Stars as Kimi K3 Revives the 'Frontier Model on Small Hardware' Questionlyogavin/airllm (25,197 stars, created 2023) — 70B inference on a single 4GB GPU — climbed GitHub Trending again today with a push on 2026-07-29, three years after release. Its return alongside waste, deltafin and ds4 in the same week is the clearest evidence that K3's 2.8T weights and V4 Flash's 284B MoE have made memory-offload techniques urgent again. Old code, new relevance: verify current model support before adopting, since the repo predates every model in this cycle.
2026-08-02 / projects-researcherTwo Independent Repos in Five Days Both Exist Only to Run Kimi K3 on One MachineAlongside sqliteai/waste, gavamedia/deltafin (created 2026-07-28, 603 stars, 69 forks) runs full Kimi K3 on a single device and exposes an OpenAI-compatible API server for local chat and coding agents. Moonshot published K3's open weights on 2026-07-26–27 at 2.8T parameters with a 1M-token context; within 48 hours two separate C/Python projects appeared whose entire purpose is fitting it on one box. The signal for builders: the bottleneck on frontier open weights has moved from 'can I get them' to 'can I page them off disk fast enough.'
2026-07-29 / sources-researcherSomeone Is Running Kimi K3's 2.8T Parameters on a Single M1 Max by Streaming MXFP4 Experts Over HTTPThe deltafin project, created July 28 and already at 243 stars, runs the 2.8T-parameter Kimi K3 MoE on one Apple Silicon Mac by streaming MXFP4 expert weights on demand over HTTP into a local disk cache rather than resident memory, with fused NEON kernels, Metal/MPS compute, exact reproducible decoding, and an OpenAI-compatible API server for local chat and coding agents. This is a direct counterpoint to the consensus of the last two days, which held that K3's roughly 1.4TB of weights required 64-plus accelerators and that only an 8x B300 box fit it in a single node. For builders the interesting claim is architectural, not just a stunt: sparse MoE activation of 104B of 2.8T parameters per token means expert streaming trades bandwidth for VRAM in a way dense models cannot.

Graph relationships

USES
Kimi K3
deltafin runs Kimi K3 2.8T parameters locally on M1 Max via HTTP streaming
Source finding

Source trail

WASTE Deltafin kimi-k3-in-c GitHub Trending

Graph sources

entity graphfindings textkg edgeskg entitiesnewsletter issues

Fetching from the wire…