Edge0 OPEN-WEIGHT · Edge0 Community Model License 1.0 + Qwen Community License (personal, research, internal use OK; commercial model services / assistants need a separate Edge0 license) · Context: 256K tokens (KV int8, 32K resident)

Edge0-145B (latent MoE)

Open-weight 145B agent model that rebuilds Qwen3.8-Flash-Next with smaller latent experts so an engine can keep hot experts on one consumer GPU and stream the rest from RAM. Edge0 reports 121.2 tok/s (Q3_K) / 98.3 tok/s (NVFP4) agent decode on an RTX 5090 D, 86.5 on Terminal-Bench 2.1 and 64.6 on DeepSWE 1.1 (vendor-reported).

Provider
Edge0 (Edge0 Community Model License 1.0 + Qwen Community License (personal, research, internal use OK; commercial model services / assistants need a separate Edge0 license))
Architecture
Latent sparse MoE derived from Qwen3.8-Flash-Next: 48 layers (hybrid linear/indexed attention), 512 experts per layer, 10 routed + 1 shared active, expert width compressed 2560→1792; 51.2B fp8 n-gram embedding table; MTP head; native 4-bit training
Parameters
145B Total / 10.12B Active
Context Window
256K tokens (KV int8, 32K resident) (Max Output: 32,768 tokens)
API Pricing
Input: $0.00 / 1M · Cached: $0.00 / 1M · Output: $0.00 / 1M
VRAM (16K Ctx)
Q4_K_M: 89.31 GB · Q8_0: 156.03 GB · FP16: 292.47 GB
Hardware Target
RTX 5090 / 4090 24 GB + 32–40 GB system RAM (expert streaming); 16 GB VRAM + 20 GB RAM tested at ~45 tok/s; Apple Silicon (Q3_K, Metal)

Interactive Token Cost Estimator (Edge0-145B)

Est. API Cost (0% Cache): --/mo | With 70% Prompt Cache: --/mo
Compare all 12 models →

GPU Hardware Compatibility Matrix (Edge0-145B at 16K Context)

GPU / Hardware Profile VRAM Q4_K_M (4.8-bit) Q8_0 (8-bit) Est. Speed (Q4)
RTX 5060 Ti 16GB
448 GB/s
16 GB REQUIRES MULTI-GPU (89.31 GB) REQUIRES MULTI-GPU (156.03 GB) ~10 tok/s
RTX 4090 24GB
1008 GB/s
24 GB REQUIRES MULTI-GPU (89.31 GB) REQUIRES MULTI-GPU (156.03 GB) ~23 tok/s
RTX 5090 32GB
1792 GB/s
32 GB REQUIRES MULTI-GPU (89.31 GB) REQUIRES MULTI-GPU (156.03 GB) ~41 tok/s
M4 Max 64GB
546 GB/s
52 GB REQUIRES MULTI-GPU (89.31 GB) REQUIRES MULTI-GPU (156.03 GB) ~13 tok/s
2× 24GB (48GB)
920 GB/s
48 GB REQUIRES MULTI-GPU (89.31 GB) REQUIRES MULTI-GPU (156.03 GB) ~21 tok/s
H100 80GB
3350 GB/s
80 GB REQUIRES MULTI-GPU (89.31 GB) REQUIRES MULTI-GPU (156.03 GB) ~77 tok/s
M4 Ultra 192GB
819 GB/s
164 GB FITS IN VRAM (89.31 GB) TIGHT FIT (Q8 KV) (156.03 GB) ~88 tok/s
8× H200 (1,128GB)
4800 GB/s
1128 GB FITS IN VRAM (89.31 GB) FITS IN VRAM (156.03 GB) ~518 tok/s

Architectural Strengths & Target Workloads

  • 145B-class quality at ~10B active parameters, designed for single-GPU expert streaming
  • Vendor-reported 86.5 Terminal-Bench 2.1 and 64.6 DeepSWE 1.1, ahead of its Qwen base
  • OpenAI- and Anthropic-compatible local server; vision encoder included

Primary Engineering Workloads: Local Agentic Coding · Private / Offline Executor Model · Single-Workstation Inference.

Frequently Asked Questions (Edge0-145B)

How much does Edge0-145B (latent MoE) cost per 1M tokens?

Edge0-145B (latent MoE) is priced at $0.00 per 1M uncached input tokens, $0.00 per 1M cached input tokens, and $0.00 per 1M output tokens (Self-hosted only via edge0-infer (OpenAI- and Anthropic-compatible endpoints) or the Edge0 Chat Windows app; no hosted inference provider at launch).

What is the context window of Edge0-145B (latent MoE)?

It supports up to 256K tokens (KV int8, 32K resident) (262,144 tokens) and a maximum generation output of 32,768 tokens.

Full TechPulse Engineering Analysis

Edge0-145B: An Open-Weight 145B Agent Model That Runs on One 24 GB GPU by Streaming Its Experts

Edge0 rebuilt Qwen3.8-Flash-Next's mixture-of-experts layer so each expert works in a smaller latent space and takes up a fixed 1.63 MB. Its engine keeps the experts it uses most on the GPU and streams the rest from system RAM. On an RTX 5090 D, Edge0 reports 98 to 121 tok/s for agent decoding and 86.5 on Terminal-Bench 2.1. The license limits commercial use.

Read Full Technical Dispatch (5 min read) →
Risograph-style illustration of a large pale grid of small tiles on a wall with about ten tiles lit in ochre, and thin ink lines running from a few of them to a small terracotta cabinet in the foreground that holds four ochre tiles