Video Learning Hub
Master multimodal AI concepts through comprehensive tutorials, guides, and best practices from our expert team.
Trusted by engineers at
Done watching? Point Mixpeek at your own video, image, audio, and document storage and make it searchable: free.
Search your own data
The Mixpeek Appliance: an AI Data Center That Starts With Four Hard Drives
The Mixpeek Appliance is a private mini data center built in phases, and the order is the design: storage first, because the data outlives every machine it will ever live on. Phase 0 is a 24-bay hot-swap chassis with four 24 TB enterprise SATA disks: 96 TB raw, roughly $5.4K for the storage server and its first disks (a planning range, per the page), on the exact hardware that later slides into the rack as-is. ZFS and an S3-compatible object service run on top, with an independent backup alongside, so the canonical s3:// namespace is real from day one. The V1 rack plan adds an RTX PRO 6000 Blackwell 96 GB, a 32-core Threadripper PRO, 512 GB ECC, and 100 GbE fabric, with the pool growing toward 288 TB raw in the same chassis; the full local Mixpeek stack runs beside vLLM so search over owned footage returns a timestamp inside a video without leaving the rack. Built for archives, defense, healthcare, financial services and manufacturing: corpora with a real reason they stay home. Scaling is addition, and the namespace never changes. Full build plan: mixpeek.com/datacenter
MVS Puts Its Write-Ahead Log on Object Storage
MVS, Mixpeek's vector store, acknowledges writes into an append-only write-ahead log and ships sealed segments to object storage as sequenced objects (wal/namespace/shard/sequence), roughly one every 20 seconds under production load. The bucket's eleven nines of durability become the durability promise. Recovery replays the chain on top of the last snapshot, replicas tail it, and change data capture reads it again: one log, three readers. Coordination is the differentiated part. Only one writer may append, every ship checks a leader lease and the writer's boot identity (a compare-and-swap fencing epoch is in flight), and garbage collection deletes a segment only when its writes are provably inside a committed snapshot, with counters that make every refusal loud instead of silent. The measured ship rate is 0.05 per second per chain against provider conditional-write ceilings of 5 to 98 per second: about 100x headroom, with chains already partitioned per namespace and shard. Learn more: mixpeek.com/mvs and docs.mixpeek.com/vector-store/overview
RL Changes 1-3% of Tokens, and They Were Already in the Top 5
A new arXiv paper (2605.06241) makes a precise mechanical claim about why reinforcement learning improves LLM reasoning: it is sparse policy selection, not capability learning. After RL training, only 1-3% of token positions change, and at those positions the promoted token always lies within the base model's top-5 alternatives. The model already contained the solution; training moved probability onto it. The authors' RL-free method, ReasonMaxxer, applies a contrastive loss only at high-entropy fork positions where the model hesitates, trains on tens of problems in minutes of single-GPU time, and matches RL outcomes at roughly three orders of magnitude lower training cost, holding across three model families, six scales, and six math reasoning benchmarks. If selection is this cheap, the expensive part of RL was the search, and knowing where the forks are is most of the work. Paper: arxiv.org/abs/2605.06241. More research we track: mixpeek.com/research
7 Gemini Flash Generations, 143 Product Photos, and a Tie Worth Trusting
We ran seven generations of Gemini Flash over the same 143 real e-commerce product images: write the caption, tags and product type, scored against each merchant's own product title with Mixpeek's F2 metric (verified identical to the server's F2AtK on all 999 predictions). The headline is a near tie: the F2 spread across all seven models is 0.035 (0.207 to 0.241), too narrow to crown a winner. What does separate: Gemini 3.7 Flash leads recall at 0.291, the axis that decides whether an asset is findable at all, and Gemini 3.5 Flash Lite names the exact product type 61% of the time against the newest full model's 42%, at 1502ms versus 3063ms per image with the set's best precision. Generation number is a weak predictor: 2.5 Flash still beats three newer models. On the embedding side only one Gemini generation is scoreable on video retrieval: gemini-embedding-001 is text-only, the -2-preview id returns byte-identical vectors to GA (cosine 1.000000), and gemini-embedding-2 scores NDCG@10 0.764 on our public 20-video benchmark. Repo: github.com/mixpeek/video-embedding-benchmark
BDH-CQ: Reasoning That Never Thinks Out Loud, at $0.0007 a Task
Pathway posted BDH-CQ to arXiv on August 10, 2026: a 150M-parameter reasoning model that reports 29.5% pass@2 on ARC-AGI-1 at a computed inference cost of $0.0007 per task, which the paper says breaks the benchmark's published cost-accuracy Pareto frontier. The mechanism is where the reasoning happens. Chain-of-thought models verbalize every intermediate step, and each of those tokens is metered compute. BDH-CQ updates a recurrent memory as examples arrive at inference time, then solves the query through iterative computation in latent space without verbalizing intermediate reasoning, so there is almost nothing to meter. It builds on Dragon Hatchling (BDH), Pathway's post-transformer recurrent architecture in which neuron-like units communicate through low-rank interactions and context lives in an evolving associative state. No frontier-model comparison numbers appear in the video: circulating GPT-5.6 figures are self-reported and unverified by ARC Prize at time of writing, so only the paper's own numbers are quoted. The paper: arxiv.org/abs/2608.09888
Meta's Muse Glimmer: a 30B Agent That Runs on One Consumer GPU
Meta open sourced Muse Glimmer on August 10, 2026: a 30 billion parameter agentic model under Apache 2.0. The figure that changes things is memory. At full precision the model wants over 55 GB; quantized it fits under 20 GB, which is one consumer GPU. Agentic here means the model closes the loop itself: it calls a function, reads what came back, and retries when the result is wrong. That retry loop is exactly what has been expensive, because every attempt was a metered round trip to a hosted endpoint, and a local model meters nothing. Meta reports speculative decoding speedups of 3.1x on an RTX 5090, 1.8x on an M5 Max and 1.5x on an M4 Max. The model reads multimodal input and speaks more than 100 languages. Weights are on Hugging Face at launch, with Ollama, LM Studio and vLLM support on day one and llama.cpp, MLX and ExecuTorch integrations coming. One thing the release does not change: a local agent still has to retrieve over something, and the corpus it grounds on has to live somewhere it can search. Meta's announcement: research.meta.ai/blog/introducing-muse-glimmer-open-agentic-model
Contextual Page Signals: Scored Entities, Brand Safety, and IAB Categories from a URL
A contextual video-advertising platform needed to know what a page is about before placing an ad on it. Point the pipeline at an article URL and it returns IAB category to tier two, entities with salience scores so the main subject reads off the structure rather than the headline, article sentiment with a confidence, a brand-safety flag with its reasons, and keywords. The brand-safety pair is the argument: two celebrity-death tributes, one flagged high_risk for death and drugs, the other safe, because the model reads the page instead of matching the word. Explore: mixpeek.com/diagrams/contextual-page-signals
21 Object Storage Providers, and the One Thing None of Them Do
We put 21 object storage providers side by side: five hyperscalers, eleven cheaper alternatives, three edge networks, one decentralized, one you run yourself. Storage price is the boring dimension, with the whole field between $6 and $25.50 per TB per month, roughly a 4x spread. Egress has no ratio at all: Google Cloud charges $120 per TB to hand your own data back, while Cloudflare R2 charges nothing, so walking out with 10 TB costs $1,200 or $0 depending only on where the bytes sit. Migration is also less painful than it looks, because 19 of the 21 expose a full or native S3 API, which makes moving mostly an endpoint change. Then there is the thing no bucket does: a list call returns keys, sizes and timestamps and says nothing about what is inside the file, so it cannot tell you which of 40,000 clips has someone holding a red umbrella. Mixpeek connects the bucket you already have, reads the contents across video, images, PDFs and audio, and makes them searchable in place. Prices re-verified against each vendor's own pricing page in July 2026. Full comparison: storage.mixpeek.com
Quantisation Scrambles What Is Near What
Generative recommendation turns items into discrete tokens so a model can generate the next item the way a language model generates the next word, and making those tokens means quantising each item embedding into a short code. In the pretrained embedding space similar items sit near each other, and that adjacency is the signal; after quantisation the intrinsic structure is disrupted, so the model's perception of item similarity is distorted before training begins. TopoTok (RecSys 2026) measures that distortion and repairs it with multi level distillation, coarse to fine: inter group for global cluster relations, intra group for local structure, inter item for fine grained alignment, reporting up to 9.42% better Recall@5. One distinction matters: this is not index-time vector compression, where you rescore the shortlist at full precision to recover accuracy. Here the structure is gone before the model trains, so there is nothing to rescore. Paper: arxiv.org/abs/2607.18600