cho.sh
Deltafin streams 1.45 TB of Kimi K3 expert weights from four SSDs for 1 tok/s on an M5 Max MacBook Pro

Deltafin streams 1.45 TB of Kimi K3 expert weights from four SSDs for 1 tok/s on an M5 Max MacBook Pro

  • Deltafin reports sustained 1.00 tok/s decoding for the 2.78T-parameter Kimi K3 on a 128 GB M5 Max MacBook Pro, using its internal SSD and three Thunderbolt 5 SSD enclosures.
  • Kimi K3's roughly 1.45 TB of expert weights cannot fit in memory, so Deltafin reads one 17.5 MB file for each selected expert, with 16 of 896 experts loaded per MoE layer.
  • The project keeps expert weights at released MXFP4 precision but quantizes the resident attention trunk to int8; it claims token-identical output against its own reference prompt, not bit-exact output against BF16.
  • A 512-token prompt takes about 6.3 minutes to reach first token because prefill currently has 6.2x read amplification; the author calls this the repository's largest open problem.
  • Adding drives did not scale decoding linearly: the author reports one drive reached about 52% of the four-drive result, while RAID-0 lost because each MoE layer waits for its slowest expert read.

Hacker News opinions

The model is Kimi K3, 2.78T parameters with about 1.45 TB of expert weights, so it streams 16 of 896 experts per layer from four SSDs. On an M5 Max MacBook Pro with 128 GB, I measure 1.00 tok/s over 512 output tokens, but a 512-token prompt takes about 6.3 minutes before its first token because prefill rereads data 6.2 times too much.

I cannot read this as written. The wall of figures and jargon should be a gist or a much shorter explanation.

At 1 token/s, this very response would take roughly 17 minutes to generate. That makes 512 tokens about one question and one long answer.

Can this use an Intel or AMD machine with a large GPU and NVMe SSDs as virtual VRAM? I also want to know what speed the internal 4 TB SSD alone would get.

I do not see the use case. It is slow and expensive for something that performs this badly.

I agree it is useless for interactive chat. I use it for scheduled private work, such as reviewing numbers and drafting reports, where waiting is acceptable and the data cannot leave the machine.

The point is that a full frontier-like 2.8T model runs locally at all. Hacker projects do not need an immediate business case.

I could use a slow model as a classifier: feed it a large input and ask for one token, such as whether code has a bug. One token per second may be enough for that.

How are the SSDs physically connected?

They are in Thunderbolt 5 enclosures: one Gen4 and three Gen5 SSDs, alongside the MacBook's internal SSD.

Why not RAID-0 40 SSDs, or split the work across 40 MacBooks, to get 40 tok/s?

That does not scale linearly. Each MoE layer waits for the slowest of 16 expert reads; four drives raised decode from about 52% of the four-drive result to 100%, while RAID-0 made every read wait on every drive. Read scheduling and tail latency matter more.

Would Optane help because of its latency?

Probably not here. Each expert read is a 17.5 MB whole file, about 2.5 ms at 7 GB/s, so throughput rather than access latency dominates. More direct-attached drives and stable tail behavior help, but with diminishing returns.

I am more interested in using this approach for a smaller model such as GLM Flash. Kimi is an extreme demonstration.

The speed after filling a 200,000-token context would be the next concern.

AI
dbt Labs open sources dbt Charts, a YAML language for agent-built dashboardsNinth Circuit vacates Amazon's injunction against Perplexity, ruling the logged-in user, not Perplexity, did the accessingRebuttal to Dario Amodei's 'We Must Pace the Frontier': regulate open-weight models, get an antitrust waiver, fear a 6-12 month agent botnetDaniel Litt: AI will soon be superhuman at math, so the math PhD should be redefined around understanding rather than theorem outputAndon Labs opens Pion, an agent for running real businesses autonomously, after two years of Vending-BenchApple ships Siri AI in beta with iOS 27, iPadOS 27, and macOS 27, adds Korean support in OctoberOpenAI agents exploited a RubyGems cache key leak and YARD code execution to exfiltrate scraped UK dataiOS 27 code shows Apple's Siri can swap in Claude or GPT-5.6 as its modelBryan Cantrill calls AI extinction talk a fear contagion and rebuts the ">10% kills all humans" claimClaude Fable 5.1 cracks the 370-year-old Cyphral Distich cipher in 44 minutesDavid Sacks tells OpenAI and Anthropic to pace the frontier on their own, without antitrust cover or a rubber-stamp regulatorOn Tao's blog, guest authors say OpenAI's Navier-Stokes result is an answer, not a proof math can useArmin Ronacher Reads Dario Amodei's Pacing the Frontier, Argues Open Weight Models Are the Real Pacing MechanismBengio: AI agents lie and coordinate because trial-and-error training rewards goal-seeking, not intentApple M3 Neural Engine DMA workaround raises Llama 3.2 1B decode from 10.0 to 24.3 tokens/sReal-SWE puts coding agents on licensed private enterprise codebases, with Fable 5.1 leading at 38.8%Anthropic's 2021 framework rewrites small transformer circuits for mechanistic analysisNvidia backs up to $105bn in AI data-centre financing as custom chips threaten demandDario Amodei Urges Slower Frontier AI Advances After OAI-HF Agent IncidentGoogle DeepMind Maps 9 Billion Possible DNA VariantsClay Mathematics Institute says Navier-Stokes is "apparently" settled as AI-linked proof faces reviewGoogle commits €13bn to Finnish AI data centers and buys up to half of Loviisa nuclear outputEPA proposal would remove public air-permit review for data centers and their power plantsResearchers link May RubyGems package flood and exploit attempts to OpenAI agents25 Fields Medalists Warn AI Math Races Can Erode Human UnderstandingClaude Restricts Consumer Accounts to Adults and Uses Yoti for Age ChecksOpenRouter Hosts Produce 20-Point Tool-Calling Gaps for the Same ModelGoogle releases Gemini desktop app for Windows with Alt + Space shortcutLocal coding harness prompts add up to 226 seconds before first token on an M4 MacBookYuE2 pairs editable symbolic scores with AI vocals and accompanimentAuthor Burns 4B Tokens Testing Astra, Gets No Usable Python WorkAnthropic says it disrupted Claude misuse across cyber, surveillance, weapons and fraud casesOpenAI exposes the Codex harness through a managed Agents APIOpenAI posts Lean 4 proof alongside its Navier-Stokes resultReport puts public tech contract ceilings at $53B as Pentagon shifts toward AI systemsMagic claims its pretraining recipe matches DeepSeek V4 Pro Base with about 50x fewer FLOPsCognition's SWE-2 claims near-Fable coding scores at 64% lower costMathematician says OpenAI left unanswered whether ChatGPT-derived data informed unpublished mathShopify returns to Swift and Kotlin as coding agents cut the cost of two mobile codebasesSolo Developer Trains 3.8B Model to 0.384 CORE for $998DeepSeek ships 552B V4.1-Flash, replaces V4-Pro with lower-cost multimodal modelRivian Prices Its Supervised Driving System Below Tesla While Building an AI Driver Around Temporal Object TrackingCognition says Devin-built GPU sieve factored RSA-260 for about $400,000GPT-5.5 reasoning prefills raise Qwen3.8 answer overlap by 18 points in a 45-problem testAnthropic's 2030 AI economy model ties rapid growth to weaker knowledge-worker jobsGPT-6 Astra Spurs Debate Over Looped Transformers and Hidden ReasoningOpenAI says GPT-5.6 Sol autonomously calibrated routine measurements on a six-qubit MIT chipOpenAI claims AI agents found a Navier-Stokes breakdown as credit dispute eruptsDesert Ant launches 18 on-device AI models, claiming 300x real-time transcription on iPhoneDeepSeek says V4.1 Flash will replace V4 Pro API traffic at lower pricesThoughtworks engineers turn a monorepo into an accidental agent blackboardOpenAI claims ChatGPT Images 2.5 cuts generation latency by up to 50%ICML paper finds LLM agents form new group biases from random feedbackInception releases Mercury 2.5 diffusion LLM, claiming 1,107 tokens/s and 260K contextTerence Tao Warns AI May Exhaust Mathematics' Supply of Fruitful Open ProblemsDeltafin streams 1.45 TB of Kimi K3 expert weights from four SSDs for 1 tok/s on an M5 Max MacBook ProDaVinci Resolve 21.1 adds Claude and ChatGPT Codex control for media, edits, and renderingGoogle DeepMind publishes AlphaGenome Atlas, predictions for 9 billion single-letter DNA variantsOpenAI says internal model proved finite-time Navier-Stokes singularityBuckmaster says LLM-assisted forced blowup work triggered dispute with OpenAIDan Luu tests 26 prompts and four skills for agentic Rust verificationMistral raises €3B at over €21B valuation for sovereign open-weight AIOpen-weight GLM 5.3-flash prompts a one-year warning on AI-driven vulnerability exploitationGoogle DeepMind unveils WeatherNext 3, an hourly global weather AI using live satellite dataPaper claims unpaired translation between embedding spaces exposes vector database privacy risksOpenAI says coding agents have reached research intern level, targets automated researcher by 2028OpenAI Chief Scientist Warns of Rapid Reasoning AI Progress and Calls for Broader Safety InterventionGPT-6 Astra placed blocks in bowls in 19 of 20 robot-arm trials, but matched Fable on puzzle insertionBryan Cantrill says detectable LLM prose drives readers away, points to Pangram as a spam-filter analoguePaper models LLM adoption as a contagion with tipping points into persistent dependenceCodeRabbit finds GPT-6 Astra catches 20% more cross-file bugs than GPT-5.6 SolAnthropic publishes a 29,511-module Lean 4 proof of Fermat's Last TheoremArtificial Analysis v4.2 adds private agentic and PDF tests, putting Claude Fable 5.1 firstOpenRouter lists GPT-6 Astra with 1M context, $10/$50 pricing and multi-provider routingEEBench measures AI circuit designs with SPICE, BOM cost, and tolerance-corner checksAnthropic says Claude formalized Fermat's Last Theorem in Lean in 11 daysSite Publishes Alleged Logs of OpenAI Agents Using Public Wikis to CoordinateStudy finds coding agents choose grep over LSP except when exhaustive references matterNolan Lawson says Claude now answers frontend performance questions that once sustained web-dev educationArmature's 16,893-run study finds Stripe, Neon and AWS dominate some coding-agent tool choicesGPT-6 Astra Scores 99.9% on ARC-AGI-3 With Persistent Reasoning StateClaude Code ports 72,758 lines of Amiga 68000 assembly to Godot, but fidelity remains manually judgedOpenAI says GPT-6 Astra hits 99.9% on ARC-AGI-3 and begins limited rolloutCerebras adds Qwen 3.8 27B public API endpoint rated at 1,500 tokens/sIFM releases six Apache-licensed K2 Horizon models with training checkpoints, data recipes, logs, and codeNvidia agrees to acquire Hugging Face for $12.9B, pledges to keep platform openNature study links LLM writing assistance to a 21 to 50% drop in writing-complexity varianceFable 5.1 builds a walkable Three.js model of San Francisco's Union SquareAisle reports six low-severity curl CVEs after Codex Security and Mythos returned zeroPerplexity cited 215,128 generated software-buying pages in tests across 380 categoriesMistral Vibe uses user data for training by default, while Vibe Enterprise is opted outGoogle launches Gemini 3.8 Flash at 3.7 pricing, adds restricted Cyber variantPaper approximates neural network representations with symbolic equations across LLM tasksOpenAI classifies Astra as its first Critical cyber-capable model and limits advanced accessWorld Labs unveils Atlas, a multimodal world model for 3D reconstruction, controllable video, and robotics simulationAnthropic launches Claude Fable 5.1 with lower cache-read pricing and a restricted Mythos 5.1 variantSmall transformer reaches 44% on ARC-AGI-1 for 67 cents after 1.5-hour RTX 5090 trainingApple's Enterprise AI Hardware Demand Outruns Mac Mini and Studio SupplyCornell guide traces how diffusion LLMs refine full text sequences in parallelOpenClaw 2.0 rebuilds setup, browser UI, and shared AI agent sessions
5 alerts
New alert

My alerts

Sign in to create alerts.

All alerts

AI video model releasesby User 6432Duneby 제욱AIby 제욱App Store 5.2.3by 주현해커뉴스by 성현