Lomash Kumar

breaking down open-source AI agents and dev tools before your feed does. videos at PandaMakingMoney. 👇

[Lomash Kumar @LomashKumar52]

MiMo-V2.6 just dropped — @XiaomiMiMo 's new trillion-parameter AI model, trained live in public, and it's completely free to download.

Xiaomi's MiMo team spent six months in silence, then opened a public dashboard and streamed the actual reinforcement learning training of two new models in real time — reward curves, cost counters, GPU failures, all visible to the world. The result is MiMo-V2.6, a full open-source model family led by MiMo-V2.6 Pro (1 trillion+ parameters) and MiMo-V2.6 Flash, alongside a speed-optimized Pro-UltraSpeed variant and a smaller Distill-Qwen-9B model for lighter hardware. In this breakdown, we cover the entire timeline of the live training run, the mixture-of-experts architecture powering these models, the native omnimodal design handling text, image, video, and audio in one system, and the benchmark numbers putting MiMo-V2.6 in direct competition with Claude Opus 5 and GPT-5.6 Sol.

If you're trying to figure out whether this open-source model is worth your attention, or how it stacks up against the current field of frontier AI models, this video walks through everything — the real API pricing, how to actually deploy it, where the benchmarks hold up, and where a few numbers don't quite add up. Whether you're a developer looking for a genuine Claude Code alternative or just tracking where open-source AI is headed in 2026, this is the full picture in one place.

[Lomash Kumar @LomashKumar52]

@opencode Letsss gooo

[Lomash Kumar @LomashKumar52]

@elonmusk Yess, it is indeed...

[Lomash Kumar @LomashKumar52]

@ajs6888 @StepFun_ai Yeahh,, agreed!!

[Lomash Kumar @LomashKumar52]

StepFun just dropped Step 5 Preview, a 600B parameter AI model with a 1M token context window and pricing that undercuts almost every model in its class.

This breakdown goes deep into what @StepFun_ai actually launched on September 20th, starting with the sparse mixture of experts architecture that keeps only 27B parameters active per token, the 92-layer narrow-deep transformer design, and the attention mechanism built to make a one million token context window actually usable. From there, we look at the pricing, including a verbosity issue that changes the real cost story, before comparing StepFun's own self-reported benchmarks against independent third-party numbers from @ArtificialAnlys, covering everything from @terminalbench to finance-specific evaluations against GPT-6 Astra and Claude Opus 5.

If you follow open source AI models, Chinese AI labs, or you're evaluating a Claude Code alternative or cheaper API option for agentic workflows, this video breaks down exactly where Step 5 Preview earns its price tag and where the open weights situation is more complicated than the announcement lets on. We also dig into the @huggingface repository StepFun already published, why it doesn't actually contain any weights yet, and what that means if you're waiting to self-host this model.

[Lomash Kumar @LomashKumar52]

Jev just got 7 free open source clones in under 48 hours, and this breaks down every single one. @typesafeai's closed, waitlist-only decision model triggered a wave of independent projects racing to build a free alternative, and this video covers all seven in detail.

You'll get a full breakdown of Kev by @jaredpalmer, Bespoke Nimble from @bespokelabsai and @madiator, jevlike by @vinnylarouge, mini-jev, jeff by @LoganMarkewich, SemIf by @theoleecj, and Laya by @Nandakishorm1 — each built on a different open model and using a different technique to approximate Jev's speed and behavior without ever seeing its actual training method. This covers exactly what base model each one runs on, how it stacks up against Jev's own published benchmarks, what hardware you'd actually need to run it yourself, and where each project is honest about falling short. If you've been searching for a Jev alternative, an open source Jev, or a free replacement for TypeSafe's decision model, this is the most complete side-by-side comparison put together so far, including the Laya vs Jev credit dispute that's been quietly brewing since launch.

This is for anyone building AI agents, working on classification pipelines, or just trying to cut API costs on repetitive decision-making tasks without sacrificing accuracy. Skip the waitlist, skip the per-token pricing, and see which of these seven free open source models is actually worth running on your own hardware today.

[Lomash Kumar @LomashKumar52]

@ajs6888 @typesafeai That's very reasonable too!!

[Lomash Kumar @LomashKumar52]

The open source model quietly destroying Jev? Laya just went head to head with a $40M funded AI startup — and the numbers are wild.

@typesafeai 's Jev launched with a former OpenAI researcher, forty million dollars in funding, and a bold claim: a new category of "System 1" model that skips text generation entirely and just returns fast, calibrated decisions. Then, one day later, an independent developer [@Nandakishorm1 ] released Laya, an open source model claiming to do the same thing for free, and the benchmark comparison table he posted looked like a complete knockout. In this breakdown, we go through what Jev actually is, how Laya's architecture works under the hood, and why the benchmark numbers being shared online are not quite as simple as they look once you read Laya's own model documentation.

This video is for anyone following the open source AI space, evaluating fast inference alternatives to large language models, or trying to figure out if Laya is actually a legitimate Jev alternative or just well-timed hype. We dig into the zero-shot accuracy numbers Laya's own team disclosed, the real-time commit history showing claims being corrected within a day of launch, and the independent Rust and Clojure reimplementations that appeared within seventy two hours, before landing on an honest verdict about which model actually deserves the attention.

[Lomash Kumar @LomashKumar52]

@CompleteSkeptic LOL 😂

[Diogo Almeida @CompleteSkeptic]

everyone: jev jev jev

my grandmother:

[Lomash Kumar @LomashKumar52]

@yevloop @typesafeai same. if it actually stays cheap and fast, agent loops feel like the obvious first win

[Lomash Kumar @LomashKumar52]

Jev just launched, and it's not another chatbot — it's a decision-making model built to replace how software makes tiny judgment calls, and the use cases are already stacking up fast.

@typesafeai new model, Jev, skips conversation entirely and focuses on one thing: fast, calibrated, typed decisions instead of generated text. In this breakdown, we cover what a "System One Model" actually is, why decision-making AI is suddenly becoming its own category separate from chat models like Claude and GPT, and the real Jev use cases already in production — from LangChain's AutoModeMiddleware guardrail system to real-time agent loops, high-volume email triage, and large-scale unstructured data processing. We also get into the training method behind it, called RLCD, and why calibrated confidence scores matter more than raw accuracy when you're trying to automate something safely.

This one's for developers, founders, and anyone building with AI agents who keeps hearing that "reasoning models" are the future and wants to understand why a non-conversational, decision-first model might matter just as much. If you're evaluating whether Jev, or this entire category of system one AI, actually belongs in your stack, or if you're just trying to keep up with what's real versus what's hype in this space, this breakdown walks through both the promise and the parts of TypeSafe's own claims that don't fully hold up under outside scrutiny yet.

[Lomash Kumar @LomashKumar52]

@TechMDAI 4 DGX Sparks cost fortune 🤣🤣

[Lomash Kumar @LomashKumar52]

@sunsetsyntax @inco_ai Same quant, same Mac, if Splash’s RSS is actually lower that’s the story, not the kernels.

[Lomash Kumar @LomashKumar52]

Splash just launched for Apple Silicon, and it's claiming to run Qwen 3.8 27B twice as fast as anything else on a Mac.

This is a full breakdown of Splash, the new open source local inference engine from @inco_ai , built specifically around two models: Qwen 3.8 27B and Qwen 3.6 35B-A3B. Instead of trying to run every model like Ollama or MLX, Splash was engineered around these two models alone, using a dedicated draft model for speculative decoding, hand tuned Metal kernels compiled for exact model shapes, and automatic memory planning based on your Mac's available unified memory. We walk through the install process, the API and coding agent integration with tools like Claude Code, OpenCode, Codex, and Hermes Agent, and break down every benchmark number Inco has published so far against Ollama, oMLX, and uzu.

If you run local AI models on a Mac, or you're deciding whether Qwen 3.8 27B is worth setting up for coding agent work, this breakdown covers what Splash actually does well, what it locks you into, and whether the "2x faster" claim holds up to real scrutiny. We also get into the bigger picture: @lmstudio 's day zero support for Splash, and what this launch says about Inco AI's broader inference business.

@Apple @Alibaba_Qwen @opencode @zhijianliu_ @jianchen1799

[Lomash Kumar @LomashKumar52]

@FReza1984 Agreed. Once the assistant can search your reading history instead of guessing, it stops being a chatbot and starts being useful.

[Lomash Kumar @LomashKumar52]

This self-hosted search engine remembers everything you've read, indexes every page and file, and even connects to your AI assistant through MCP.

Hister is a free, open source, self-hosted search engine built by the same developer behind Searx, designed to solve one specific problem: finding something you already read but can no longer locate. Instead of crawling the entire internet like Google or Bing, Hister builds a private full-text index out of the web pages you visit and the local files you keep, covering PDFs, Word documents, Markdown, and plain text, then lets you search through the actual content using a real query language with filters, wildcards, negation, and date ranges. This breakdown covers what Hister actually is, the origin story behind its creator, a full walkthrough of its core features including offline previews and semantic search, how its Model Context Protocol server lets an AI assistant search your personal archive directly, a step-by-step look at installation, and an honest privacy deep dive into how your data is actually stored and handled.

If you're someone who reads constantly across dozens of tabs, saves documents locally, or is already building a personal AI setup and wants your assistant to have real memory instead of guesswork, this is exactly the kind of self-hosted infrastructure worth understanding. Whether you're comparing it to browser bookmark managers, tools like Linkding or Karakeep, or just looking for the best way to build a personal knowledge base for AI in 2026, this video walks through everything you need before deciding if it belongs in your own setup.

@GithubProjects @github

[Lomash Kumar @LomashKumar52]

@harisahmad59 Jev is the trending topic for now!!!

[Lomash Kumar @LomashKumar52]

@cocktailpeanut LOL... That would crazzyyy

[Lomash Kumar @LomashKumar52]

@axhiii88547 Bro served us well!!!

[Lomash Kumar @LomashKumar52]

@TKNetX @inco_ai Yeah that model-specific route is kinda wild. Specializing instead of being general-purpose actually paid off.

[Lomash Kumar @LomashKumar52]

@meimu7ns4 @inco_ai Yeah that’s the real question. Fast on a 30-second bench is one thing. Fast after 40 minutes of agent loops is another.