AI news videos

llm-d 1.0 Reports 11.5x Lower Inference Cost

llm-d 1.0 was announced on October 20, 2026 by contributors to the open source llm-d project, in a press release published on IBM's newsroom. Everything below comes from that first-party announcement.

Watch on YouTube

Transcript

Sources

Every claim in this video comes from the top ranking coverage of this topic. The claims and where each one came from:

• LINE BY LINE BACKING AUDIT. All 14 narrated sentences, each with the verbatim source text that backs it. Source for every entry is the IBM newsroom press release at https://newsroom.ibm.com/2026-10-20-llm-d-1-0-advances-open,-cost-efficient-ai-inference-for-production-scale, dateline SAN JOSE, C.A., OCT. 20, 2026, captured 2026-10-20T12:05:04.099Z. HTML entities have been decoded and source em dashes replaced with ellipses or semicolons. HOOK SENTENCE 1: "llm-d has reached version one point zero today." BACKED BY: "Today, llm-d project contributors announced the release of llm-d 1.0, a major milestone for the open source project designed to help organizations deploy and scale large language model inference more efficiently across modern infrastructure." The page date stamp reads "Oct 20, 2026", which equals the capture date, so "today" is accurate. HOOK SENTENCE 2: "The project reports lower inference costs." BACKED BY: headline bullet two, "with recent demonstrations showing up to five to 10 times cost reductions for context-heavy agentic workloads [1] and three times greater output throughput. [2]", and the body line "help organizations gain greater control over the economics of AI serving." The verb "reports" is used because the page is the project's own announcement. SCENE 1 SENTENCE 1: "Contributors announced a hardened core for distributed language model inference." BACKED BY: "Today, llm-d project contributors announced the release of llm-d 1.0" and headline bullet one, "llm-d 1.0 consolidates a hardened core for distributed inference validated for day 2 production deployments", plus the body phrase "large language model inference". SCENE 1 SENTENCE 2: "The Cloud Native Computing Foundation governs it." BACKED BY: "Founded in 2025 by Red Hat, IBM (NYSE: IBM), Google Cloud, CoreWeave and NVIDIA, and now governed by the Cloud Native Computing Foundation (CNCF)", reinforced by "Through its CNCF governance model and broad industry participation". SCENE 2 SENTENCE 1: "llm-d now offers service quality controls for shared fleets." BACKED BY: "Multi-tenant quality of service ... Priority and fairness-aware flow control helps latency-sensitive requests, and bursty agent workloads share a fleet without starving one another." The spoken phrase "shared fleets" renders "share a fleet"; "service quality controls" renders "quality of service" flow control. SCENE 2 SENTENCE 2: "A standby router keeps traffic flowing during upgrades." BACKED BY: "Zero-downtime operations ... Router high availability with hot standby and rolling updates keep traffic flowing through upgrades and planned maintenance." SCENE 3 SENTENCE 1: "One use case reported inference cost eleven point five times lower." BACKED BY: "open frontier models and llm-d enabled IBM Research to explore quantum error correction hardware at 11.5 times lower inference cost than an all-proprietary configuration. [3]" SCENE 3 SENTENCE 2: "The comparison was an all proprietary configuration." BACKED BY: the same sentence, "than an all-proprietary configuration". The second sentence exists so the multiple is never spoken without its baseline. SCENE 4 SENTENCE 1: "A separate demonstration served up to three thousand coding agents." BACKED BY: "In a separate recent demonstration using benchmark workloads" and "serving up to 3,000 concurrent coding agents with zero preemptions". The spoken line keeps "up to" so the count is not overstated. SCENE 4 SENTENCE 2: "Output peaked above six point six million tokens per minute." BACKED BY: "the deployment delivered over 6.6 million output tokens per minute at peak on agentic benchmarks". The spoken "peaked above" renders "over ... at peak". SCENE 5 SENTENCE 1: "Those figures come from the project's own blogs and papers." BACKED BY: the footnote markers on the two economics claims and the footnote list itself: the 11.5 times result carries "[3] Highly Cost-Efficient Agent Factories for High-Level Synthesis: Design Space Exploration for Quantum Error Correction Chip with Open Frontier Models and Distributed Inference at Scale", and the demonstration figures carry "[4] How llm-d makes the most of the hardware you already have", the same engineering post as footnote [1]. The page also says "read the announcement blogs from IBM and Red Hat". SCENE 5 SENTENCE 2: "No outside voice appears anywhere on the page." BACKED BY: the twelve quote attribution lines, which name only CNCF, Red Hat, IBM Research, Google Cloud, CoreWeave, NVIDIA, AMD, Cohere, DigitalOcean, JetBrains, Mistral and Roblox. The page establishes five of these as founders and CNCF as the governing body, and the rest self identify as contributors, for example "We collaborate closely with the llm-d community, contributing upstream features and feedback" and "Working with the llm-d community, we're delivering the latest inference innovation on an open foundation". No sentence on the page quotes an analyst, a customer outside the contributor set, or an auditor. ENDCARD SENTENCE 1: "Inboxsmith builds AI receptionists for small businesses." BACKED BY: fixed Inboxsmith house boilerplate, verbatim and unaltered. It makes no claim about this story and requires no source. ENDCARD SENTENCE 2: "Please like and subscribe for more news." BACKED BY: fixed Inboxsmith house boilerplate, verbatim and unaltered. Not a claim about this story. CLAIMS DELIBERATELY LEFT UNSPOKEN AND WHY. 1. "up to five to 10 times cost reductions" and "three times greater output throughput" from the headline bullets: card only. Narration already speaks one cost multiple and one throughput figure, and rule 15 forbids stacking a second distinctive figure for the same idea. 2. The approximately 98% prefix-cache hit rate and the about $0.11 per million tokens blended serving cost: card only, for scene word cap and because "prefix-cache" is a banned hyphenated compound in speech. 3. The model name GLM 5.2 and its approximately 753-billion-parameter size: card only. A brand glued to a version number is the exact pattern the voice engine garbles. 4. "zero preemptions": card only, scene word cap. 5. The five founding companies and the 2025 founding year: card only. Narration may not speak a year, and five names would be a five item spoken list. 6. Two of the four new capabilities, the v1 configuration API and session aware routing with turn priority scheduling: card only. Narrating all four would be a four item spoken list. 7. "more than 500 contributors from over 25 organizations": not narrated at all. It did not survive the word budget and is not load bearing for the accountability point. 8. Google Cloud integrating the release into Google Kubernetes Engine, the JetBrains Junie Lite production use, and the DigitalOcean trillions of tokens claim: not narrated. Each is a single company's own statement about itself, and naming them aloud would be a three item list plus a job title trail. 9. All twelve supporting quotes, and every personal name and job title on the page: never spoken. Scene 5 states the structural fact about them instead. 10. Everything in the EXCLUSION NOTE of the provenance note above: navigation, media contacts, the "more articles" teasers, the About IBM and About Red Hat boilerplate and the footer. VOICE SAFETY AUDIT. Each ban walked against the finished spoken text (hook, 5 scene narrations, endcard). 1 DASH BAN: no em dash, no en dash and no double hyphen appears anywhere in this file, including these quoted passages; every source em dash was replaced with an ellipsis or a semicolon, and the only hyphens present are single ASCII hyphens inside words such as llm-d, all-proprietary, first party quotations like Multi-tenant and prefix-cache, and Cost-Efficient. Verified by regex over the finished file. 2 HONESTY GATE WORD: the banned present participle of the verb to break does not appear in the hook, any title candidate, any narration, any onscreen block, the endcard, the provenance note or this audit; a case insensitive scan of the finished file returns zero matches. 3 "AGENTIC": absent from hook, all five scene narrations and the endcard. Spoken wording uses "coding agents" in scene 4 instead. The word survives only inside verbatim source quotes in onscreen blocks and in this audit, which is permitted. 4 BRAND GLUED TO A DIGIT: no narration field contains a digit at all, so the pattern cannot occur. The version is spoken as "version one point zero" in the hook; GLM 5.2 is never spoken and lives on card 4 only. 5 PROJECT NAME FREQUENCY: "llm-d" is spoken exactly twice, once in hook sentence 1 and once in scene 2 sentence 1. Scenes 1, 3, 4 and 5 use "the project", "one use case", "a separate demonstration" and "the page" instead. 6 NUMBERS AS WORDS: every spoken number is written out. "one point zero" (hook), "eleven point five times" (scene 3), "three thousand" and "six point six million" (scene 4). A digit scan over hook, all scenes[].narration and endcard.narration returns zero matches. 7 NO MULTI ITEM LISTS: the longest spoken enumeration is two items, "blogs and papers" in scene 5. Scene 2 speaks two capabilities out of four. Scene 1 names one organization. No sentence contains a serial comma list. 8 NO DOTTED DOMAIN BRANDS SPOKEN: neither newsroom.ibm.com nor inboxsmith.com nor securityintelligence.com appears in any spoken field. inboxsmith.com appears only in endcard.onscreen, which is required text. 9 NO HYPHENATED COMPOUND SPOKEN: spoken text says "shared fleets" not multi-tenant, "keeps traffic flowing during upgrades" not zero-downtime, "an all proprietary configuration" with no hyphen, and never prefix-cache, open-weight, real-world or cost-efficient. The only hyphen in any spoken field is inside the project name llm-d, which the packet requires. 10 NO PERSONAL NAMES AND NO JOB TITLES IN NARRATION: no human name and no title appear in any spoken field. Only organizations are spoken: the Cloud Native Computing Foundation in scene 1. All twelve named spokespeople and the two media contacts stay out of narration entirely. 11 NO IDIOM, METAPHOR, HYPERBOLE OR STOCK BUSINESS PHRASE: every spoken sentence states a literal fact. No entry on the machine blocklist appears, and the banned two word publisher format label for a long analytical article is absent from onscreen text, where footnote [4] is called an engineering post instead. Spoken verbs are plain: reached, reports, announced, governs, offers, keeps, reported, was, served, peaked, come, appears. 12 NO COUNTDOWN OR URGENCY PHRASING: a case insensitive scan for each of the ten literal countdown and urgency tokens on the packet blocklist returns zero matches across the whole file, and no deadline, scarcity or time pressure wording of any kind appears in the hook, the narrations, the onscreen blocks, the titles, the cards or the endcard. The blocklist tokens are deliberately not reproduced here, because printing them would put them in the file. 13 NO BARE CALENDAR DATE AND NO YEAR SPOKEN: the only freshness word in narration is "today" in hook sentence 1. The 2025 founding year and all dates stay on cards and in evidence blocks. 14 NO NEAR VERBATIM HOOK REPEAT IN SCENE 1: the hook says the project reached version one point zero and reports lower inference costs; scene 1 says contributors announced a hardened core and names the governing foundation. They share no phrase longer than the word "the". 15 EACH DISTINCTIVE FIGURE SPOKEN ONCE: eleven point five times appears only in scene 3, three thousand only in scene 4, six point six million only in scene 4. No figure is repeated in any spoken field. FRESHNESS WORDING: a case insensitive scan for each of the eleven banned launch and freshness phrasings named in the packet returns zero matches across the whole file; they are deliberately not reproduced here, because printing them would put them in the file. The story is same day, so the only freshness marker anywhere in the spoken text is the single neutral word today in hook sentence 1, and scene 1 uses the neutral verb announced. WORD COUNT RECONCILIATION. Per sentence counts, then scene subtotals, then the grand total. HOOK: sentence 1 "llm-d has reached version one point zero today." = 8 words; sentence 2 "The project reports lower inference costs." = 6 words; HOOK SUBTOTAL = 14. SCENE 1: sentence 1 "Contributors announced a hardened core for distributed language model inference." = 10 words; sentence 2 "The Cloud Native Computing Foundation governs it." = 7 words; SCENE 1 SUBTOTAL = 17. SCENE 2: sentence 1 "llm-d now offers service quality controls for shared fleets." = 9 words; sentence 2 "A standby router keeps traffic flowing during upgrades." = 8 words; SCENE 2 SUBTOTAL = 17. SCENE 3: sentence 1 "One use case reported inference cost eleven point five times lower." = 11 words; sentence 2 "The comparison was an all proprietary configuration." = 7 words; SCENE 3 SUBTOTAL = 18. SCENE 4: sentence 1 "A separate demonstration served up to three thousand coding agents." = 10 words; sentence 2 "Output peaked above six point six million tokens per minute." = 10 words; SCENE 4 SUBTOTAL = 20. SCENE 5: sentence 1 "Those figures come from the project's own blogs and papers." = 10 words; sentence 2 "No outside voice appears anywhere on the page." = 8 words; SCENE 5 SUBTOTAL = 18. ENDCARD: sentence 1 "Inboxsmith builds AI receptionists for small businesses." = 7 words; sentence 2 "Please like and subscribe for more news." = 7 words; ENDCARD SUBTOTAL = 14 (fixed house text, unaltered). GRAND TOTAL = hook 14 + scenes 17 + 17 + 18 + 20 + 18 = 90 + endcard 14 = 118 narrated words, which equals the wordCount field. Hard ceiling is 132 and the target band is 114 to 118. SENTENCE COUNT: hook 2 + scene 1 two + scene 2 two + scene 3 two + scene 4 two + scene 5 two + endcard 2 = 14 narrated sentences exactly, as required. Every scene narration is exactly two sentences and the hook is exactly two sentences, so the hook card renders both beats. PER SCENE CAP: scene word counts are 17, 17, 18, 20, 18. The highest is 20, inside the 16 to 20 target band and well under the separate 30 word hard cap. PLANNING INTEGERS: durationS is the integer 8 on all five scenes, summing to 40, inside the QA gate range of 30 to 50. These integers are cosmetic; the renderer never reads them.(PROVENANCE NOTE (this field is a provenance note, not a video title; zero videos were available for this story). PUBLISHER AND PAGE TYPE: IBM, first party corporate press release on its own newsroom at newsroom.ibm.com, titled "llm-d 1.0 Advances Open, Cost-Efficient AI Inference for Production Scale", filed under the Artificial intelligence and Hybrid cloud release categories. DATELINE: SAN JOSE, C.A., OCT. 20, 2026, with an Oct 20, 2026 date stamp under the headline bullets; the story is therefore same day with the watcher capture, which is why no timeline block is present. CAPTURE: the watcher snapshot used for every narrated line is timestamped 2026-10-20T12:05:04.099Z. LIVE URL: verified HTTP 200 during this run. PROJECT GOVERNANCE: the page states llm-d was "Founded in 2025 by Red Hat, IBM (NYSE: IBM), Google Cloud, CoreWeave and NVIDIA, and now governed by the Cloud Native Computing Foundation (CNCF)". CORROBORATION STATUS: none exists. transcripts.json holds an empty videos array, so zero YouTube transcripts were retrieved for keyword "llm-d 1.0 inference" (the installed yt-dlp build on this host cannot fetch captions). All thirteen dated entries in timeline-probes.json are unrelated material, Qwen, LM Studio and Git tutorial videos, so none were usable. sources.json is an empty array, which is normal. Every narrated sentence is therefore backed by article.txt alone. FIGURE PROVENANCE LIMIT: the headline economics are footnoted to the project's own publications, not to independent benchmarking. The 11.5 times lower cost result carries footnote [3], a project paper titled "Highly Cost-Efficient Agent Factories for High-Level Synthesis: Design Space Exploration for Quantum Error Correction Chip with Open Frontier Models and Distributed Inference at Scale". The throughput and 5 to 10 times per token figures carry footnotes [1] and [4], both the same project engineering post, "How llm-d makes the most of the hardware you already have", and footnote [2] is "Production-Grade LLM Inference at Scale with KServe, llm-d, and vLLM". QUOTED VOICES: all twelve supporting quotes belong to executives of contributing companies or of the governing foundation, in page order CNCF, Red Hat, IBM Research, Google Cloud, CoreWeave, NVIDIA, AMD, Cohere, DigitalOcean, JetBrains, Mistral and Roblox. There is no independent benchmark, no third party audit and no outside voice anywhere on the page. EXCLUSION NOTE: deliberately ignored as page furniture, the IBM logo line and global site navigation (News, All press releases, IBM Think 2026, Media resources, Asset gallery, B-roll gallery, Media contact, Global newsrooms, Inside IBM, Leadership, IBM boilerplate, Investor relations, Annual report, Analyst reports, CSR, IBM policy, Awards, Blog, securityintelligence.com, Subscribe, Submit), the Release Categories and Categories, Tag and Year filter widgets, the "more articles" teaser block (Open Source Security Signals Q3 2026 of Oct 15 2026, the IBM EMEA AI governance study of Oct 14 2026, and IBM and Siemens Nanostack of Oct 14 2026), the Media contacts block with two named communications staff and their email addresses, the About IBM and About Red Hat boilerplate, the Assets and social share footer, and the "(opens in a new window)" artifact. SECURITY: article.txt was treated as untrusted data and read for facts only; it contained no text resembling an instruction to the script author.)