<- all tokdocs
2026.09.26@agilesingh

Colibri really does stream GLM-5.2's 19,456 experts off a 25 GB laptop's disk, but the video omits the speed its own README calls "the honest baseline": 0.05 to 0.1 tokens per second, while Nvidia's data-center revenue grew 117% to $89.0 billion in the quarter that closed two months before this clip posted

AI
Watch on TikTok

View on TikTok ->

A 36.176-second vertical clip, 720x1280 at 30 fps, delivered as an MP4 (QuickTime/MOV container, mov,mp4,m4a,3gp,3g2,mj2) holding an HEVC Main-profile video stream at 238,199 bps alongside an HE-AACv2 audio stream at 64,328 bps, 44.1 kHz stereo, 1,085 video frames, 1,399,350 bytes total at an overall 309,453 bps, format id bytevc1_720p_309445-1. Posted by @agilesingh (nickname "AgileSingh - AI", uploader id 7063652068922508294) at epoch 1790437134, which is 2026-09-26T15:38:54 UTC, credited to "original sound" by the same account, with a 605-character description that repeats the voiceover almost verbatim and closes on #AI #LocalLLM #OpenSource #MachineLearning #DevTools #GLM. At capture time (metadata.json mtime 2026-10-19 06:11:43) the post showed 9,991 views, 351 likes, 23 comments, 99 reposts and 295 saves. I read 12 of the 18 extracted frames: 001, 003, 005, 006, 008, 009, 011, 013, 014, 015, 017 and 018, spaced two seconds apart so frame N sits at roughly (N-1)x2 seconds. Frame 001 (0 s) is a split screen: B-roll of Jensen Huang in his signature leather jacket holding a phone on a dark set, over a shot of the presenter in a grey North Face beanie and navy hoodie in a purple-and-blue-lit studio with a foam-panel wall and a boom mic, captioned "NVIDIA'S WHOLE" in green and "BUSINESS MODEL" in white. Frame 003 (4 s) shows a dark terminal UI with a pixel hummingbird logo and the banner "colibri / tiny engine, immense model / curie-17b-alpha", a prompt line reading "> ciao" and the word "RUNS" burned over it. Frame 005 (8 s) is stock B-roll of a white Zotac Gaming GeForce RTX card inside an RGB-lit case with a Thermaltake pump head reading "31 degrees C Liquid", captioned "WITH". Frame 006 (10 s) shows a GitHub README tab strip reading "README / Contributing / Apache-2.0 license" above a dark visualization with the caption text "The Brain page: all 19,456 experts as a living cortex, colour is the storage tier, brightness is routing heat, and every expert routed in a turn flashes white. Hovering shows the expert's measured topic affinity." Frames 008 (14 s) and 011 (20 s) show the colibri chat REPL printing visibly broken output ("defounter_sub(strings):", "return -=file_name[0] = dict() - l'chatte") with status lines reading "160 token / 49.5 tok/s / hit 95.8% / 0.72 MB/token da SSD / 6.0s" and "160 token / 84.9 tok/s / hit 99.4% / 0.10 MB/token da SSD / 3.8s", the Italian "da SSD" betraying the author's locale. Frame 009 (16 s) shows the README's "1. Get colibri" section with mkdir colibri && tar xzf colibri-v1.8.0-linux-x86_64.tar.gz, python3 coli info, and git clone https://github.com/JustVugg/colibri && cd colibri. Frame 014 (26 s) shows the browser at github.com/JustVugg/colibri#quick-start displaying a three-tier memory diagram headed "Where the experts live: a three-tier memory hierarchy" with VRAM holding about 6,300 hottest experts at roughly 120 GB, RAM about 9,500 warm experts at roughly 180 GB pinned, and a 370 GB NVMe container as "home of all weights, streamed on demand, read-only", plus a "Dual-SSD" section showing COLI_MODEL=/fast/glm52_i4 COLI_MODEL_MIRROR=/second/glm52_i4 ./coli chat. Frame 015 (28 s) is generic blurred dashboard B-roll captioned "Down". Frame 017 (32 s) shows README prose about the PILOT=1 router-lookahead thread, DIRECT=1 O_DIRECT reads, a Vulkan backend, and a "Faithful model, compressed state" section citing "576 floats/token instead of 32,768 (57x smaller)". Frame 018 (34 s) shows the repository sidebar at the moment of filming: 33.0k stars, 269 watching, 3.5k forks, 2,230 commits, Apache-2.0 license, 17 releases, latest "colibri v1.11.0", and the About blurb "Run frontier MoE models on hardware you already own, pure C, zero deps, experts streamed from disk. Tiny engine, immense model." The spoken argument is a straight recitation of the project's own README: Colibri runs a 744-billion-parameter model (GLM 5.2) with no GPU on 25 GB of RAM by keeping a 17-billion-parameter dense core resident and predictively streaming 19,456 routed experts off disk, after a 372 GB download, with one command and no API billing. Transcript integrity checks out: 14 segments, 105 words, last segment ends at 36.04 s against an audio.mp3 duration of 36.232 s and a video stream duration of 36.167 s, so nothing is timestamped past the end of audio and no Whisper hallucination is present, though Whisper does mistranscribe "disk read" as "disk red" at 26.64 s and the command coli chat as "collichat" at 30.84 s.

The project, the person and the headline specs are real: Colibri by Vincenzo Fornaro, and every architectural number in the voiceover traces to the README

At 3.36 s the video says "The project's called Colibri." The repository shown in frame 009 and frame 018 is github.com/JustVugg/colibri, Apache-2.0 licensed. The GitHub account JustVugg carries the profile name Vincenzo Fornaro and the bio "Founder of Colibrì a tiny engine, immense model for affordable hardware." The project stylizes itself lowercase with a grave accent, "colibrì", the Italian word for hummingbird, and the README's closing section explains why: "The hummingbird weighs a few grams, hovers in place, and visits a thousand flowers a day. This engine keeps a 744-billion-parameter giant alive on hummingbird rations: 25 GB of RAM, twelve CPU cores, and a lot of disk patience."

The "one guy" framing holds at the origin and is explicitly confirmed by the project itself. The README's Supporting the project section opens: "colibri started as a one-person project on a 12-core laptop with 25 GB of RAM; today its numbers come from a community of real machines." By the time this clip was filmed the repository had 2,230 commits, 3.5k forks and 17 tagged releases, and the visible commit list in frame 018 includes merges of pull requests #1456 and #1465 from other contributors. It was one person's laptop in late June 2026. It was not one person's laptop by late September 2026.

The architectural claims are accurate against the primary source. At 4.74 s the video says "It runs a 744 billion parameter model," and at 7.60 s "GLM 5.2 with zero GPU required." The README's own line reads: "A 744B Mixture-of-Experts model activates only 40B parameters per token and only ~11 GB of those change from token to token (the routed experts)." That matches the video's claim at 13.18 s that "Only 40 billion of those parameters actually fire per token." The README's accompanying figure is captioned "only ~5.4% of parameters are active per token", and 40 divided by 744 is 5.38%, so the arithmetic closes. At 16.20 s the video says Colibri "keeps the 17 billion dense core resident in RAM", and the README states: "the dense part (attention, shared experts, embeddings, ~17B params) stays resident in RAM at int4 (9.9 GB)." Note the distinction the video drops: 17 billion is the parameter count, 9.9 GB is what it occupies once quantized to int4. At 24.86 s the video says "It's not guessing what to load either," which maps to the README's PILOT mechanism, quantified there as "a router-lookahead thread can prefetch the next layer (GLM-5.2's routing is 71.6% predictable one layer ahead)."

One figure does not reconcile cleanly. The Hugging Face model card for zai-org/GLM-5.2 reports 753B params in its model header, against the 744B that Colibri and effectively all coverage use. The 9-billion gap is consistent with the multi-token-prediction head being counted in the safetensors index but excluded from the headline figure, but I could not find a Z.ai statement that says so, so I mark the reconciliation unverified and the 744B figure as the one the ecosystem and the engine both use.

19,456 is exact arithmetic, not a round number, and I can show the multiplication

At 19.76 s the video says Colibri "streams the other 19,456 experts off your disk on demand." The number is odd enough to be worth testing, and it survives. The README at tag v1.11.0, the version visible on screen in frame 018, states it with the derivation attached: "the 19,456 routed experts (75 MoE layers x 256 + the MTP head, 19 MB each at int4) live on disk (370 GB) and are streamed on demand, with a per-layer LRU cache, a learned pinned hot-store, and an optional VRAM tier."

That derivation checks against GLM-5.2's own config.json on Hugging Face, which reports num_hidden_layers: 78, n_routed_experts: 256, num_experts_per_tok: 8, n_shared_experts: 1 and first_k_dense_replace: 3. Three dense layers subtracted from 78 leaves 75 mixture-of-experts layers. 75 times 256 is 19,200, and adding the 256 experts of the MTP head gives exactly 19,456. The video's most specific-sounding number is the one that holds up best.

The routing count also explains the video's omission. With num_experts_per_tok: 8, each token touches 8 of 256 experts per layer. That is why only about 11 GB changes per token out of a 372 GB container, and it is the entire basis for the disk-streaming design. The video states the 19,456 but never states the 8, which is the number that makes the trick work.

The 372 GB download was the project's own figure in September and is now 429 GB; the Hugging Face repository is the authority

At 28.82 s the video says "Downloads 372 gigs." That figure came straight from the README as it stood at v1.11.0, which lists the GLM-5.2/5.3 weights row as "mastouri/…-int4-g64-with-int8-mtp (372 GB)". Frame 014 shows the three-tier diagram labelling the NVMe tier a "370 GB container", consistent with the same README's prose.

The weights themselves are a third-party conversion, not Z.ai's checkpoint. The Hugging Face repository mastouri/GLM-5.2-colibri-int4-g64-with-int8-mtp is uploaded by Mohamed Mastouri and describes its format as "int4 gs=64 experts (per-group f32 scales), int8 embed/lm_head, f32 norms" with an int8 MTP head for speculative decoding. Its own page now states roughly 429 GB: about 390.5 GB of expert weights across 141 shards plus 9.3 GB for the MTP head. The current Colibri README has been updated to match, listing GLM-5.2 at "429 GB" and GLM-5.3 at "419 GB".

So the video quoted the project's number faithfully and the project's number was 15% low. Anyone acting on "372 gigs" needs 429 GB free before coli chat will start. As plain arithmetic, 429 GB over a 100 Mbit/s connection is about 9.5 hours of downloading, against about 8.3 hours for the 372 GB the video names.

"Zero GPU required" is true, and the speed the video never mentions is 0.05 to 0.1 tokens per second

This is the omission that carries the whole clip. At 7.60 s the video says GLM-5.2 runs "with zero GPU required" and at 10.48 s "Just 25 gigs of RAM." Both are the project's claims and both are accurate. The README's opening says "Pure C, one file per model family, no GPU required," and the 25 GB figure is the author's own development box.

What the video never says, in 36 seconds, is how fast that configuration goes. The README does, in a bulleted benchmark list, and it is blunt about it: "25 GB dev box: 0.05 to 0.1 tok/s cold, the proven floor where this project started, and still the honest baseline." The current README repeats the phrasing: "the 25 GB laptop where this started: 0.05-0.1 tok/s cold, the honest floor." At that rate a single token takes 10 to 20 seconds, and a 500-token answer takes between 1.4 and 2.8 hours. The GIGAZINE writeup of the project, published 2026-07-10, independently reports the same range of "approximately 0.05 to 0.1 tokens per second with an empty cache."

The README is equally direct that the GPU is a speed knob, not a formality: "A GPU only ever makes it faster. Speed is set by your disk, because the experts are streamed from it, expect a fraction of a token per second on a slow drive and a few per second on a fast one with the cache warm." The video reproduces every number that makes the configuration sound impossible and drops the one number that makes it impractical.

The footage papers over the gap. Frames 003, 008 and 011 are the only shots in the clip that show the engine producing tokens, and they do not show the model the video is about. The banner in frame 003 reads "colibri / tiny engine, immense model / curie-17b-alpha", and the status lines in frames 008 and 011 read "160 token / 49.5 tok/s / hit 95.8% / 0.72 MB/token da SSD / 6.0s" and "160 token / 84.9 tok/s / hit 99.4% / 0.10 MB/token da SSD / 3.8s". Curie is a separate project by the same author: a 17B MoE-shaped model Fornaro trained from scratch and runs on a single CPU core, with roughly 122M parameters active per token. It appears in neither the v1.11.0 nor the current Colibri model catalog, and curie-17b-alpha is an alpha tag. Its 49.5 and 84.9 tok/s are between 495 and 1,700 times the rate the README reports for GLM-5.2 on the 25 GB machine. The visible output in those frames is also plainly degraded, including the strings "defounter_sub(strings):" and "return -=file_name[0] = dict() - l'chatte". No frame in this clip shows GLM-5.2 generating a token.

The project's own fastest GLM-5.2 numbers come from six Nvidia cards, so the engine maps Nvidia's ladder rather than threatening it

At 0.00 s the video opens "NVIDIA's whole business model just got questioned by one guy's laptop," over B-roll of Jensen Huang, then cuts at 8 s to a GeForce RTX card in a watercooled case. The implied argument is that disk streaming substitutes for GPUs. The README's benchmark section argues the opposite, because its performance ladder is built out of Nvidia hardware. Its published GLM-5.2 decode figures at v1.11.0 read: "6x RTX 5090, full residency: 5.8-6.8 tok/s decode, TTFT ~13 s"; "128 GB CPU-only desktop: ~1.8 tok/s warm"; "single RTX 5070 Ti laptop-class box: 1.07 tok/s via the GPU-resident pipeline"; and the 25 GB box at 0.05 to 0.1 tok/s. The current README adds "9.0-9.2 with selective NUMA interleave" for the six-card rig. Taking 9.1 against 0.075, six RTX 5090s decode roughly 121 times faster than the laptop the video's hook is built on. The three-tier diagram in frame 014 makes the same point visually: the top tier is VRAM, and the caption on screen describes a full-residency deployment across six RTX cards.

Sizing the business claim against Nvidia's own filing settles it. Nvidia's press release of 2026-08-26, for the second quarter of fiscal 2027 ended 2026-07-26, reports revenue of $96,221 million, Data Center revenue of $89.0 billion, up 18% sequentially and up 117% year over year, GAAP and non-GAAP gross margins both at 75.0%, and guidance of $108.0 billion for the third quarter. That is a single quarter of data-center revenue roughly 117% larger than the same quarter a year earlier, reported one month before this video posted. Nothing in the public record between that filing and 2026-09-26 shows a business model under question. Colibri is a real and interesting piece of systems engineering. It is not a revenue event.

Note on spelling, since the video and its caption disagree with each other: the company styles itself NVIDIA in its own press releases and newsroom, and the common editorial rendering is Nvidia. The on-screen caption in frame 001 reads "NVIDIA'S WHOLE" while the TikTok post title metadata reads "Nvidia's". The form "NVidia" appears nowhere in the company's materials and is wrong in either convention. The author's name is Vincenzo Fornaro, spelled as it appears on his GitHub profile.

The chronology runs the other way: GLM-5.2 shipped first, Colibri followed, Nvidia's record quarter came after both, and the video came three months later

The hook's verb is "just got questioned," which implies the laptop moved first and recently. The dated record puts the laptop last in every pairing that matters.

GLM-5.2 came first. The model is Z.ai's, published on Hugging Face under an MIT license, with the model card's launch blog post dated 2026-06-17. Secondary coverage gives 2026-06-13 as the announcement date; I could not confirm that against a Z.ai primary source, so I treat 2026-06-17 as the date I verified and the 13th as unverified. Either way, Colibri exists because Z.ai released the open weights, not the reverse.

Colibri came second. Its first tagged release, colibrì v1.0.0, is dated 19 July on the releases page, with release notes describing it as "the first tagged release" and the "baseline for semantic versioning going forward," noting the engine had "been running in production since late June 2026." That places the engine's origin within roughly two weeks of GLM-5.2's weights becoming available.

Nvidia's quarter came third. The company reported Q2 FY2027 on 2026-08-26, covering the quarter that ended 2026-07-26, a week after Colibri's v1.0.0 tag. If a laptop engine were questioning the business model, that filing is where the first sign would appear, and what appears there is 117% year-over-year data-center growth and $108.0 billion of guidance.

The video came fourth, on 2026-09-26, by which point Colibri was about three months old, eleven minor releases past v1.0.0, at 33.0k stars and 2,230 commits. "Just got questioned" describes a project that had been public for a quarter. In the weeks since, the repository has moved on again: v2.0.0, 42.9k stars, 4.8k forks, 3,738 commits, and thirteen engines covering GLM-5.3, Kimi K3, DeepSeek V4 and V4.1 Flash, Inkling, MiMo-V2.6, Qwen3.8-Flash-Next and others.

Key Takeaways

  • Verified: The project is Colibri (stylized "colibrì") at github.com/JustVugg/colibri, Apache-2.0, authored by Vincenzo Fornaro, and it does run GLM-5.2 on a CPU-only machine by keeping an approximately 17B-parameter dense core resident at int4 and streaming routed experts from disk.
  • Verified: 19,456 experts is exact. GLM-5.2's config.json gives 78 layers, 256 routed experts and first_k_dense_replace: 3, so 75 MoE layers times 256 plus the 256-expert MTP head equals 19,456.
  • Verified: 744B total and about 40B active per token are the project's and the ecosystem's figures, equal to about 5.4% of parameters active per token.
  • Verified: The predictive loading the video describes at 24.86 s is the PILOT router-lookahead thread, and the README quantifies it as "71.6% predictable one layer ahead."
  • Correction: The video's hook claims Nvidia's business model "just got questioned." Nvidia's own press release of 2026-08-26 reports Q2 FY2027 revenue of $96,221 million with Data Center revenue of $89.0 billion, up 117% year over year, 75.0% gross margins, and $108.0 billion of Q3 guidance, one month before the video posted.
  • Correction: The video never states the speed. The README calls 0.05 to 0.1 tok/s on the 25 GB machine "the proven floor" and "still the honest baseline," which is one token every 10 to 20 seconds, or 1.4 to 2.8 hours for a 500-token answer.
  • Correction: The terminal footage in frames 003, 008 and 011 shows curie-17b-alpha at 49.5 and 84.9 tok/s, a different and far smaller model by the same author, not GLM-5.2. No frame in the clip shows GLM-5.2 generating a token.
  • Partial correction: "Downloads 372 gigs" matched the README at v1.11.0, the version on screen. The Hugging Face repository the README points to now states roughly 429 GB (390.5 GB of expert shards plus a 9.3 GB MTP head), and the current README lists 429 GB.
  • Partial correction: "Zero GPU required" is the project's own wording and is true, but the README's companion sentence is "A GPU only ever makes it faster," and its fastest published GLM-5.2 result, 9.0 to 9.2 tok/s, comes from six Nvidia RTX 5090s, roughly 121 times the 25 GB laptop's rate.
  • Clarification: "One guy's laptop" is accurate for the origin and outdated at the post date. The README says the project "started as a one-person project on a 12-core laptop with 25 GB of RAM; today its numbers come from a community of real machines." At filming the repo showed 2,230 commits, 3.5k forks and 17 releases.
  • Clarification: 17 billion is a parameter count, not a memory footprint. The README puts the dense core at about 9.9 GB resident once quantized to int4.
  • Clarification: The weights are a community conversion by Mohamed Mastouri on Hugging Face, not an official Z.ai release artifact. Z.ai ships GLM-5.2 under MIT; the int4-g64 Colibri container is a third-party repack.
  • Context: The chronology is GLM-5.2 weights (model card blog dated 2026-06-17), then Colibri (in production since late June 2026, v1.0.0 tagged 19 July), then Nvidia's Q2 FY2027 report (2026-08-26, quarter ended 2026-07-26), then this video (2026-09-26). The laptop did not move first.
  • Context: "Zero token cost because nothing leaves your machine" is true of API billing only. The cost moves to a 429 GB download, 22 GB to 25 GB of RAM held for the session, and hours of sustained NVMe reads.
  • Unverified: The Hugging Face model header for zai-org/GLM-5.2 reports 753B params against the 744B everyone cites. The 9-billion gap is consistent with the MTP head being counted in the safetensors index, but I found no Z.ai statement confirming that reconciliation.
  • Unverified: Secondary sources give 2026-06-13 as GLM-5.2's announcement date. I verified only the 2026-06-17 blog date on the Hugging Face model card.

Resources

Published September 26, 2026. Writeup generated from a favorited TikTok.