FOUNDING OFFERSign up to unlock 35% off your first plan.

What Are Edge0-145B and Edge0-120B? On-Device MoE Specs and Local Visual Workflows

SUMMARY

Edge0-145B and Edge0-120B are two on-device latent sparse Mixture-of-Experts (MoE) models released on November 12, 2026, rebuilt from Qwen3.8-Flash-Next for local coding, reasoning, and visual-inspection agent loops. Edge0-145B targets higher agentic coding margins with a larger active expert budget, whereas Edge0-120B trims the expert pool to lower memory residency and speed up constrained decoding. Although both models support optional vision encoding for auditing screenshots or 3D renders alongside code, those local terminal loops do not retouch product photos or build commercial campaigns. For independent businesses wanting ready-to-share social promotions rather than workstation GPU tooling, Stillmore turns everyday phone photos into finished images, videos, and captions handled by creative specialists, and its team may evaluate models such as Edge0-145B or Edge0-120B when a brief suits them while you control publishing.

Lan He

Stillmore Blog cover for Edge0-145B, Edge0-120B, featuring black faceted sphere with luminous ring.
Contents

What Are Edge0-145B and Edge0-120B?

Both models are locally executable latent sparse MoE systems engineered to run multi-step coding, reasoning, and tool-using tasks on personal workstations and compatible laptops without cloud API calls.

I reviewed Edge0's November 12 launch post, the Edge0 directory on Hugging Face, and the official model cards for Edge0-145B-NVFP4 and Edge0-120B-NVFP4. Instead of compressing an existing checkpoint after training, Edge0 rebuilt the routed expert layers from Qwen/Qwen3.8-Flash-Next, trained natively at 4 bits, and listed Q3_K and NVFP4 builds alongside its smaller 35B, 8B, speech, and embedding releases under dual community licenses.

In practical software engineering work, teams select between the two releases based on local memory capacity and latency priorities. A group debugging large code repositories on a high-memory workstation favors Edge0-145B for stronger coding margins, while a developer running tighter host memory often chooses Edge0-120B for snappier interactive replies.

Why On-Device Latent MoE Models Matter for Local Work

Local inference above 100 billion parameters runs into memory capacity and memory bandwidth ceilings long before compute saturation, which Edge0 addresses by shrinking routed expert records.

Before late 2026, most personal hardware stayed confined to sub-10B models because large weights competed with the operating system for limited memory, as discussed in Edge0's release announcement. That announcement highlights a 29GB peak memory measurement on a compatible laptop, while the Edge0-120B-NVFP4 model card explicitly clarifies that reported test budgets represent measurement conditions rather than universal minimum system requirements. Local execution avoids cloud API token fees, though compatible hardware and electricity still carry direct costs.

For an engineering team working on unreleased firmware or proprietary hardware layouts, running inference locally keeps sensitive files off external servers. That privacy benefit still assumes the team owns suitable machines and accepts the ongoing maintenance of local runtimes, storage tiers, and runtime flags.

How Latent MoE Architecture and Tiered Memory Loading Work

Replacing standard full-width experts with narrow latent projections reduces the bytes moved per generated token and allows weights to stream across GPU VRAM, system RAM, and SSD storage.

As outlined in the Edge0-145B-NVFP4 and Edge0-120B-NVFP4 documentation, shared projection layers route tokens through narrow latent experts, while disjoint caching flags prevent duplicate expert copies across GPU and host memory. The compact table below summarizes the key published figures once for direct comparison.

Specification Edge0-145B Edge0-120B
Active / total parameters 10.12B / 145B 8.64B / 120B
Layers / latent width (l) 48 / 1792 48 / 1280
Build sizes (Q3_K / NVFP4) 40.81 / 47.79 GiB 27.98 / 32.96 GiB
RTX 5090 D NVFP4 agent decode 98.3 tok/s 135.5 tok/s
Terminal-Bench 2.1 86.52 81.3 (blog) / 75.3 (card)*
Reported peak / license 29GB setup; Edge0 + Qwen 1.0 Same reported peak and license

*Note: Edge0's launch blog reports 81.3 on Terminal-Bench 2.1 for Edge0-120B, whereas the Edge0-120B-NVFP4 model card header lists 75.3.

As an illustrative deployment scenario, consider a solo developer using a 24 GB desktop GPU alongside moderate host memory who wants to inspect local codebases without upgrading RAM sticks. By reserving desktop VRAM for display stability and streaming colder experts from fast solid-state storage, the developer trades peak token throughput for fitting a 100B-tier agent within existing workstation limits.

Local Visual and Agent Workflows in Practice

Both Edge0 models execute multi-step coding, 3D scene scripting, and visual verification loops locally, with each variant trading benchmark margin against memory footprint and decode responsiveness.

In Edge0's launch post, Edge0-145B demonstrates scripting 3D city scenes in Blender, assembling an explainer video pipeline, routing a board in KiCad for Unreal Engine 5 animation, and generating indoor drone navigation code. The Edge0-145B-NVFP4 and Edge0-120B-NVFP4 guides add that reading images requires pairing mmproj-BF16.gguf with an external strata-vision helper configured in model.json, with options to keep the vision encoder on CPU or allocate VRAM.

Local Workflow Scenario Primary Bottleneck to Audit Practical Configuration Choice Operational Tradeoff
Complex repository debugging on high-RAM host Agent accuracy on long tool chains Choose Edge0-145B with higher resident expert cache Demands more host RAM and storage footprint
Iterative coding under tighter VRAM and RAM Constrained memory swap stalls Choose Edge0-120B with capped RAM and SSD offload Trades modest coding margin for smoother decode
Visual render checks alongside coding loops Display and context VRAM headroom Run strata-vision on CPU when VRAM is tight Slower image encoding in exchange for expert cache space
Customer-facing social ad production Brand accuracy, retouching, and time cost Skip local model hosting and use managed creative delivery Relies on creative specialists rather than local terminals

Consider an illustrative independent game studio validating Blender camera paths offline so unreleased environment art never leaves its local network. A technical artist can script geometry fixes and feed viewport screenshots into the local vision loop to catch clipping errors, yet that engineering audit remains entirely separate from producing retail storefront advertisements.

Raw Model Outputs vs. Finished Creative Assets

Using a local vision-capable agent to inspect a render or answer prompts about an image is fundamentally different from producing brand-accurate commercial promotions ready for social channels.

As detailed in Inside Stillmore: Our Creative Production Workflow, effective promotional content requires art direction matched to the item, compositing and retouching, motion pacing that keeps offer wording readable, and detailed review against reference photos. Because a local vision encoder converts pictures into language-model embeddings rather than operating a retouching suite (Edge0-120B-NVFP4 README), raw agent outputs do not automatically preserve exact packaging proportions, match lighting and shadows, or guarantee legible label text.

Production Checkpoint Local Edge0 Agent Output Commercial Ad Standard Stillmore Verification Focus
Source material Codebases, JSON flags, prompts, or render paths Everyday phone photos of products, spaces, or people Enlarged product outline and legible packaging labels
Creative direction User writes prompts and scripts each iteration Concept tailored to launch vs. everyday reminder Framing that highlights the selling feature cleanly
Retouching and motion Generates scripts or text feedback on screenshots Seamless compositing, shadows, and readable motion timing Photography, Photoshop, and After Effects integration
Accuracy review Developer checks every frame and text line manually Finished piece matches the actual item on sale Pre-delivery checks for shape, color, text, and motion
Delivery package Raw terminal text, scripts, or local render files Ready-to-post image or video plus matching caption Delivered to account with merchant publishing control

Imagine an illustrative neighborhood cafe introducing a layered seasonal drink for weekend foot traffic. Its poster and short reel must preserve the exact glass silhouette, drink gradient, printed cup label, and promotion dates on a small phone screen, work that relies on human art direction, compositing, typography review, and motion pacing rather than a local coding terminal.

When Edge0-145B or Edge0-120B Fits, and When It Does Not

Edge0-145B and Edge0-120B fit technical teams with supported hardware running internal engineering tasks, whereas time-pressed business owners needing finished social promotions are better served by managed creative production.

Before deploying locally, technical evaluators must check platform and license boundaries noted on the Edge0-145B-NVFP4 and Edge0-120B-NVFP4 cards: NVFP4 targets CUDA on Linux or Windows, Q3_K extends to Apple Silicon Metal, community CMF conversions are unendorsed third-party builds, and dual community licenses require separate written authorization plus upstream compliance for commercial bundling or independent commercial assistants. For independent merchants who want ongoing marketing content without configuring inference flags or auditing model terms, Stillmore handles creative direction, production, and pre-delivery checks from your uploaded phone photos; our team may evaluate and use models such as Edge0-145B or Edge0-120B when they suit a brief, combining them with editing and creative review to seek the best result while you choose when and where to publish.

As an illustrative merchant scenario, an independent apparel shop can browse Inspiration to select a specific look for a seasonal coat drop while letting automatic creation cover routine weekly posts. Saved photos stay reusable in the shop's Library as long as the product remains unchanged, avoiding local hardware upgrades or software upkeep.

How to Evaluate Local Models or Start Finished Content Production

Where you begin depends on whether your team intends to compile local inference weights for technical workflows or upload business photos to receive checked marketing assets.

Developers evaluating Edge0 locally start by selecting a single model folder from Edge0's Hugging Face directory, building edge0-strata and edge0-infer or launching the Windows desktop app, and adding vision configuration if needed per the Edge0-145B-NVFP4 and Edge0-120B-NVFP4 guides. Business owners focused on social promotions instead can follow Stillmore's introductory guide: upload clear everyday photos to Library once, pick an Inspiration direction or enable automatic creation, and receive finished images, videos, and captions.

Decision Step Local Edge0 Evaluation Track Stillmore Creative Production Track
Hardware audit Confirm GPU CUDA or Apple Silicon support, RAM headroom, and SSD bandwidth Confirm any phone or computer with existing business photos
Input preparation Clone engine repos, select model pack, and configure model.json flags Zoom in on photos to verify clear outlines, labels, and promo text
Workflow ownership Engineer writes prompts, debugs scripts, and tests visual checks Creative specialists plan, produce, edit, and review finished assets
Licensing check Review Edge0 and Qwen community terms for intended internal or product use Purchase ongoing agency service without managing model licenses
Output handoff Export and assemble downstream files in external tools Receive finished images, videos, and captions ready to publish yourself

Before pulling tens of gigabytes of weights, technical teams should verify disk space, memory headroom, and license fit against their internal roadmap. By contrast, a small business preparing a promotion only needs clear product shots and accurate offer wording before submitting a request or turning on automatic creation.

Conclusion

Edge0-145B and Edge0-120B demonstrate how rebuilding MoE experts into narrow latent projections brings 100B-class coding and visual-inspection loops onto compatible personal hardware. Choosing between them comes down to balancing Edge0-145B's higher coding margins against Edge0-120B's lighter footprint and faster constrained decoding. Even with local vision support, however, running a developer agent on a workstation remains a distinct job from delivering brand-accurate promotional images, videos, and captions for your business channels.

Resources

Source Direct Link What It Verifies
Edge0 Launch Post (Nov 12, 2026) Introducing Edge0-145B Release scope, 29GB laptop measurement context, local workflow demos, and 81.3 benchmark figure
Hugging Face Directory Edge0 Models on Hugging Face Public repository listings for Edge0-145B, Edge0-120B, and sibling model families
Edge0-145B Model Card Edge0-145B-NVFP4 README 145B latent MoE specs, throughput, memory flags, vision setup, and licensing
Edge0-120B Model Card Edge0-120B-NVFP4 README 120B latent MoE specs, test-limit clarification, 75.3 header figure, and licensing
Stillmore Workflow Guide Inside Stillmore: Our Creative Production Workflow Photography, VFX, selective AI production workflow, and pre-delivery accuracy checks

Frequently Asked Questions

Should a small business run Edge0-145B or Edge0-120B locally to create social ads?

No, most independent businesses should not run Edge0-145B or Edge0-120B locally to create social ads. Both models are local coding and reasoning engines built for compatible workstation hardware, command-line configuration, and technical iteration rather than turnkey ad production. Turning code or scene outputs into commercial promotions still requires manual editing, layout, and brand checks. Stillmore instead begins with everyday photos already on your phone and delivers finished images, videos, and captions produced by photography, post-production, and creative teams, freeing you to stay focused on customers.

How should teams weigh Edge0-120B's smaller footprint against Edge0-145B, or avoid hardware sizing altogether?

Engineering groups weigh Edge0-120B's lighter memory residency and faster constrained decoding against Edge0-145B's stronger coding margins when provisioning local GPUs or Apple Silicon hosts. If you operate a neighborhood shop rather than a software team, you can skip hardware sizing, storage offload flags, and memory budgets altogether. With Stillmore, creative specialists select and coordinate the internal production tools behind each brief, so your business receives finished promotional images, videos, and captions without buying workstation graphics cards or configuring local model folders.

Does Edge0's vision encoder edit or retouch product photos for ads?

No, Edge0's vision support encodes images so the local language model can inspect renders or answer visual questions during coding tasks. It does not retouch real-world photos, match storefront lighting, or composite social motion graphics. At Stillmore, our photography, post-production, and VFX specialists use tools such as Adobe Photoshop and After Effects alongside selective generative tools when a concept calls for them. Before delivery, the team reviews product shape, color, texture, label text, and pacing against your reference images so customers recognize what you actually sell.

What commercial licensing limits apply to Edge0-145B and Edge0-120B versus Stillmore deliverables?

Both models carry dual community licenses rather than unrestricted open-source terms, requiring separate written Edge0 authorization plus upstream Qwen compliance before commercially bundling the weights or launching an independent commercial assistant. By contrast, Stillmore clients purchase an advertising creative agency service rather than managing raw model weights or third-party inference binaries. You supply existing photos and promotion context, and Stillmore delivers finished promotional images, videos, and captions prepared specifically for your business channels without shifting model licensing administration onto your team.

Does Stillmore already use Edge0-145B or Edge0-120B in client projects?

Stillmore does not claim that Edge0-145B or Edge0-120B is already integrated into active client workflows. Instead, our AI production team may evaluate and use models such as Edge0-145B or Edge0-120B when they suit a brief, combining them with editing and creative review to seek a stronger final piece. Tool selection, agent coordination, compositing, and pre-delivery accuracy checks stay with our specialists, so merchants never have to test weights, compare new releases, or manage technical workflows themselves when planning regular social updates.

What materials do I prepare for Stillmore compared with prompting a local agent?

Running a local agent requires preparing model folders, runtime configuration files, or technical scripts and manually checking every generated output. With Stillmore, you only gather everyday phone photos where your product's outline, packaging, and printed label stay legible when enlarged, alongside accurate item names and any promotion dates. Our creative team uses those source materials to shape the visual direction, produce the finished image or video, write matching captions, and verify details against your submission before delivery so the final piece reflects your actual offering.

Can the same product photos be reused across multiple Stillmore campaigns?

Yes, Stillmore stores uploaded product, storefront, and team photos in your account Library so suitable shots can support new promotions without scheduling fresh shoots every month. Reuse depends on what the next concept needs: a clean front angle may suit a poster while the product stays unchanged, whereas an obscured side detail or updated packaging requires another photo. When new items or angles appear, you simply upload those additions rather than rebuilding a local workflow from scratch each time a promotion changes.

Can I pick specific campaign styles while automating routine updates in Stillmore?

Yes, Stillmore lets you combine self-selected requests with ongoing automatic creation inside My Creations. After uploading at least 3 suitable photos to your Library on Try it or 15 photos on Growth, you can turn on "Stillmore creates for me" for regular updates while still picking an Inspiration style whenever a launch needs specific direction. Both modes share the same Plan Credits, and switching back to manual mode stops new automatic pieces while completing active work already underway in your queue.

How does Stillmore's delivery timeline compare with local model generation?

Local inference speed still leaves setup, prompting, post-editing, and manual quality checks on your own schedule before anything is publishable. At Stillmore, requests you submit yourself are delivered within 5 days once payment, required materials, and your request are complete. On Growth in automatic mode, your first video arrives within 5 days after you upload at least 15 Library photos, enable automatic creation, and have payment, Credits, and suitable materials ready, followed normally by one video every 5 days, though late-month uploads cannot guarantee full month-end completion.

How do Stillmore Plan Credits cover images versus videos?

Stillmore Plan Credits cover both videos and images from a single balance without asking you to track GPU memory or API token consumption. Each finished video uses 3,000 Credits and each finished image uses 500 Credits, whether you select an Inspiration style yourself or let automatic creation schedule updates. Unused Plan Credits can accumulate up to two months of your plan allowance, capped at 6,000 Credits on Try it or 36,000 Credits on Growth, and purchased Add-on Credits remain available without expiring.

Does Edge0 or Stillmore post directly to Instagram, TikTok, or YouTube Shorts?

Edge0 provides local model weights and inference software without a bundled social-media publishing service, so automated posting would require building separate scripts or external integrations. Stillmore also leaves publishing decisions in your hands rather than posting directly to your social accounts. Once our team produces and checks your promotional images, videos, and captions, the finished files arrive in your Stillmore account alongside an email notification, leaving you in full control of when and where to share on Instagram, TikTok, or YouTube Shorts.

About the author

Lan He

I'm Lan. Years of writing promotional and editorial work in the US, mostly for companies with big audiences and bigger expectations — where a single article might reach hundreds of thousands of readers, and where the good ones kept earning traffic long after the campaign ended. I learned what makes a page survive that: one clear audience, one real question, images that show the actual work, and a structure that keeps people reading. Now I write for small business owners who do everything themselves, answering the questions they actually ask: what's worth posting, which photos are worth taking, and how to tell one story across images, video, and words. I've seen good content carry a small company further than its ad budget ever did.

More about Lan He & all articles ↗

Give them a reasonto stop by.

Send us your phone photos. Get posts ready to share.

See what we can make
Stillmore fashion creative look
All stories ↗