Strata runs 125-billion-parameter agent loops on system RAM

Step 5 Preview uses a Mixture-of-Experts architecture with 600 billion total parameters and 27 billion activated parameters. According to woshipm, the system ran autonomously for over 3 hours to parse 3D models while writing and debugging code for an interactive 3D antique appreciation game. Sustaining that workload requires cheap context. Step 5 Preview sets its API pricing at 7 yuan per million input tokens and 20 yuan per million output tokens, which woshipm reports is one-eighth the price of Opus 5.

Token economics now dictates agent design. Continuous autonomous loops consume vast token volumes, turning operating expenses into the primary ceiling on deployment.

Other architectures undercut standard rates even further. According to woshipm, MiMo-V2.6-Flash costs 0.02 yuan per million cached input tokens and 2 yuan per million output tokens, whereas MiMo-V2.6-Pro costs 0.025 yuan per million cached input tokens and 6 yuan per million output tokens. As ifanr observed, APPSO argues that users who only rely on cloud AI services like ChatGPT and Gemini do not yet need to replace their laptops specifically for local model execution. For continuous developer loops, that pricing divide shifts the focus from raw reasoning scores to hardware economics.

How long-running agent loops break frontier token budgets

Autonomous developer loops break traditional inference budgets through compounding token volume. In benchmark testing recorded by woshipm, Grok 4.7's speed dropped from 60 to 38.8 as generated tokens surged from 94 million to 240 million. That expansion doubled the cost per task from $1.86 to $3.74. Extended workflows multiply these totals. Woshipm documented Step 5 Preview running autonomously for over 3 hours to parse 3D models and write code for an interactive 3D antique appreciation game.

Unbounded loops make unit economics the defining bottleneck. When continuous agent runs consume hundreds of millions of tokens, standard frontier API pricing becomes unsustainable.

Sparse Mixture-of-Experts architectures mitigate this cost by activating only a small slice of their network per cycle. Step 5 Preview relies on 600 billion total parameters. Only 27 billion activated parameters engage during an active forward pass. Its API costs 7 yuan per million input tokens and 20 yuan per million output tokens, which woshipm notes is one-eighth the price of Opus 5.

Xiaomi applies an identical structural strategy. MiMo-V2.6 adopts the base architecture of MiMo-V2.5, pairing 1.02 trillion total parameters with 42 billion activated parameters. MiMo-V2.6-Pro costs 0.025 yuan per million cached input tokens and 6 yuan per million output tokens. For MiMo-V2.6-Flash, the cost drops to 0.02 yuan per million cached input tokens and 2 yuan per million output tokens. Non-real-time batch inference cuts both rates by 50%. According to ifanr, APPSO argues that users relying purely on cloud services like ChatGPT and Gemini do not yet need to replace laptops for local model execution.

Step 5 and MiMo-V2.6 bring trillion-parameter reasoning to developer tools

Local host runtimes are already reaching consumer hardware. According to ifanr, WorkBuddy on the Asus ProArt 16 RTX Spark includes a beta feature allowing users to download and manage locally running AI models directly within the application. For isolated operations, Microsoft is developing the Microsoft Execution Containers project to provide isolated execution environments for AI agents performing file operations and tool calls. Yet complex system tasks quickly exceed the ceiling of smaller local engines.

Autonomous engineering demands massive capacity. In an engineering stress test reported by woshipm, Step 5 Preview received instructions to check out the open-source linkding repository at commit f32af96123e9749e60c54caa3ac7709621faffee and generate 100,000 bookmarks to profile performance.

Step 5 Preview operates across 600 billion total parameters with 27 billion activated parameters and a 1 million token context window. In tests tracked by woshipm, this architecture ported the open-source local network streaming application Liuguang Yiying (流光逸影) from Windows to macOS and Linux in a single pass. It detected hidden bugs and updated documentation without human intervention. In another run, it produced a playable single-file HTML Canvas tower defense game with automatic pause upon window blur in a single shot. Local models cannot match this breadth.

Cloud scale delivers comparable results on standard benchmarks. According to woshipm, MiMo-V2.6-Pro scored 46.32 points on Artificial Analysis, ranking first among open-source and domestic Chinese models while tying Grok 4.7. On DeepSWE v1.1, the benchmark score of MiMo-V2.6-Pro rose from 58.4 to 72.57, while Flash rose from 48.8 to 65.68. MiMo-V2.6 handles direct video inputs to replicate websites from screen recordings. When tested by woshipm, the model even installed Blender on a computer where it was missing to finish a 3D modeling task.

Frontier Agent Execution Architectures: Subsidized Cloud MoE vs. Local Host-RAM Offloading

Strata trades dedicated GPU silicon for high-speed system memory

According to appinn, the open-source runtime Strata runs the 125-billion-parameter Qwen3.8-Flash-Next model on consumer PCs equipped with 8GB of VRAM. The uncompressed model requires 360GB of VRAM, an impossible ceiling for consumer graphics cards. Strata bypasses that limit by offloading bulk weights into system RAM and SSD storage. Because the model activates 6 billion parameters out of 125 billion total parameters per inference step, the graphics processor only handles the active slice.

Host memory requirements scale rapidly with quantization precision. Strata requires 32GB of system RAM for entry-level compressed weights, 48GB for Q2_0 and IQ2_XS, 64GB for IQ3_XXS and IQ3_S, and 96GB for IQ4 models.

Disk speed defines whether that architecture functions. When Qingxiaowa tested Strata with mechanical hard drives for appinn, performance proved impractically slow. Solid-state throughput delivers usable rates: an 8GB RTX 4060 Ti paired with 32GB of system RAM yields 19 to 21 tokens per second on the Coder IQ1_M quantization. Strata officially recommends a minimum of 12GB of VRAM across NVIDIA or AMD graphics cards and 32GB of host RAM. On macOS, the runtime handles a single concurrent request and context remains verified up to 32K.

Unified memory architectures sidestep host-bus bottlenecks but introduce severe purchase costs. According to ifanr, the Asus ProArt 16 RTX Spark laptop lists at 28,999 yuan for the 64GB unified memory edition and 37,999 yuan for the 128GB unified memory edition. Its 64GB unified LPDDR5X configuration offers approximately 46.7 GB of usable GPU memory over a theoretical peak bandwidth of 300 GB/s.

Agent Execution and Model Routing Decision Framework

• Running massive-scale background agent tasks, continuous code loops, or batch evaluations where token costs dominate and real-time generation is secondary. Route jobs to Xiaomi's MiMo-V2.6-Flash or MiMo-V2.6-Pro using non-real-time batch inference. With the 50% batch discount applied to base pricing (MiMo-V2.6-Flash at 0.02 yuan per million cached input tokens and 2 yuan per million output tokens; MiMo-V2.6-Pro at 0.025 yuan per million cached input tokens and 6 yuan per million output tokens), costs are drastically reduced for token-heavy agent cycles.
• Executing multi-hour autonomous engineering loops that require multimodal comprehension, long-horizon planning, or up to 1 million tokens of context. Deploy StepFun's Step 5 Preview via API at 7 yuan per million input tokens and 20 yuan per million output tokens (one-eighth the price of Opus 5). While its 600-billion-parameter MoE architecture handles 1-million-token contexts and complex multi-hour tasks, note that its detailed thinking process increases completion time.
• Deploying large sparse MoE models locally on consumer desktop hardware with limited VRAM (8GB to 24GB). Use the open-source Strata runtime (Niko1221/Strata on GitHub) with NVMe SSD storage and at least 32GB of system RAM (scaling to 48GB for Q2_0, 64GB for IQ3_S, and 96GB for IQ4). Avoid mechanical hard drives, which yield impractically slow performance. Expect 19 to 21 tokens per second on an 8GB RTX 4060 Ti with 32GB RAM (Coder IQ1_M), 60 to 77.9 tokens per second on a 24GB RTX 3090 with 64GB DDR4 (IQ3_S), or up to 94 tokens per second on a 12GB RTX 5070 with 64GB RAM (Q2_0).
• Executing local agent workflows entirely on a single portable workstation without offload bus latency. Use an integrated system like the Asus ProArt 16 RTX Spark (priced at 28,999 yuan for 64GB unified memory or 37,999 yuan for 128GB unified memory). The 64GB model delivers 46.7 GB of usable GPU memory at 300 GB/s bandwidth, reaching 55 to 60 tokens per second on Qwen3 32B in NVFP4, or run built-in models like the 16 GB Tencent 50-billion-parameter MoE inside WorkBuddy.
• Running local sparse MoE inference on Apple Silicon Macs. Deploy Strata v0.1.44 using its native Metal backend on macOS 15 or later for Qwen3.8-Flash-Next. Be aware that the macOS implementation is restricted to a single concurrent request, verified up to a 32K context, and does not support Qwen3.6-35B-A3B. Hardware selection is critical: a Mac mini M4 with 32GB memory reaches only 8 to 14 tokens per second on Flash-Next Q2_0, whereas a Mac M5 Max with 128GB memory achieves 95 to 103 tokens per second on Flash-Next IQ3_S.

Why zero-cost local execution clashes with agent wall-clock latency

Local inference eliminates token fees, but generation speed dictates whether autonomous loops remain practical. According to appinn, Qingxiaowa considers 40 tokens per second usable and over 60 tokens per second very smooth. Entry-level machines lag behind. In LM Studio on an Asus ProArt 16 RTX Spark, ifanr recorded Qwen 3.8 27B generating 14.3 tokens per second with a first-token latency of 2.83 seconds, alongside prompt processing at 615.7 tokens per second. Benchmarks from appinn show a Mac mini M4 with 32GB of memory running Flash-Next Q2_0 at 8 to 14 tokens per second.

Sluggish throughput compounds across multi-turn agent runs. When tasks demand dozens of tool iterations, low generation speeds turn rapid debugging into a multi-hour wait.

Strata narrows that gap by offloading execution to host RAM. On a 12GB RTX 5070 with 64GB of RAM, appinn reports that Strata hits 94 tokens per second on Q2_0 quantization and 53 tokens per second on IQ3_S. Older hardware benefits too. A 24GB RTX 3090 with 64GB of DDR4 RAM maintains 60 to 77.9 tokens per second on Qwen3.8-Flash-Next in IQ3_S. In AMD Radeon AI PRO R9700 testing with the 35B model, Strata achieved 83 tokens per second, jumping to 161 tokens per second with MTP 2 enabled. On a Mac M5 Max with 128GB of memory, Flash-Next IQ3_S ran at 95 to 103 tokens per second.

Physical hardware limits also shape offline loops. Continuous inference on a 27-billion-parameter model pulled 86.5W to 88.4W from the 99.9 Wh battery of an unplugged Asus ProArt 16 RTX Spark, according to ifanr. The 64GB version reached 55 to 60 tokens per second using Qwen3 32B in NVFP4 precision. For offline agent harness needs, WorkBuddy embeds an optimized 16 GB Tencent 50-billion-parameter mixture-of-experts model accelerated by CUDA 13.

Hybrid routing balances those local constraints against wall-clock urgency. The woshipm author considers Step 5 Preview's thinking process overly detailed, causing coding tasks to take a relatively long time. When immediate turnarounds are essential, developers bypass local wait times with fast cloud tiers. According to woshipm, MiMo-V2.6-Pro-UltraSpeed generates output up to 20 times faster than MiMo-V2.6-Pro.

Agent loops degrade quickly when token bloat outpaces execution speed. In benchmarks recorded by woshipm, Grok 4.7 saw speed drop from 60 to 38.8 as token volume climbed from 94 million to 240 million, doubling per-task expense from $1.86 to $3.74. Running long loops requires dampening that cost curve. MiMo-V2.6 routes through 42 billion activated parameters out of 1.02 trillion total, lifting DeepSWE v1.1 scores from 58.4 to 72.57 on Pro and 48.8 to 65.68 on Flash. The model also ranked third among open-source systems on Terminal-Bench 4.0, behind GLM-5.3 and Qwen 3.8 Max.

Woshipm author Karl will seriously consider promoting MiMo-V2.6 to a primary daily agent model. During testing, it automatically installed Blender on a computer where the software was missing to complete a 3D modeling task.

For readers outside China

• Availability: Strata and its AMD-compatible counterpart Gufo are open-source software projects available globally on GitHub (under Niko1221/Strata for Strata). Cloud APIs for Xiaomi (MiMo-V2.6 series) and StepFun (Step 5 Preview) are accessible online, though whether international credit cards or non-mainland phone numbers are accepted is not disclosed in sources. StepFun plans to release the model weights for Step 5 Preview on October 15. The Asus ProArt 16 RTX Spark laptop is currently quoted in Chinese yuan, with international retail distribution not disclosed in sources.
• Pricing: API token pricing is billed in Chinese yuan: Xiaomi MiMo-V2.6-Flash costs 0.02 yuan per million cached input tokens and 2 yuan per million output tokens; MiMo-V2.6-Pro costs 0.025 yuan per million cached input tokens and 6 yuan per million output tokens (both eligible for a 50% batch inference discount). StepFun Step 5 Preview API costs 7 yuan per million input tokens and 20 yuan per million output tokens. For local hardware, the Asus ProArt 16 RTX Spark laptop costs 28,999 yuan for the 64GB unified memory edition and 37,999 yuan for the 128GB unified memory edition. Open-source runtimes Strata and Gufo are free.
• Closest Western equivalents: Claude Opus 4.6 / Opus 5; Grok 4.7; ChatGPT and Gemini; llama.cpp (with Windows ML integration)
• Data residency: The source material does not cover data residency policies, server locations, GDPR compliance, or overseas API routing for Xiaomi or StepFun endpoints. For environments with strict data isolation requirements, the sources highlight entirely local options: running open weights locally via Strata, running WorkBuddy's built-in 16 GB Tencent 50-billion-parameter MoE on unified memory hardware, or isolating agents using Microsoft Execution Containers.

Sources

• woshipm 实测小米新模型MiMo V2.6，我真服了第一次看开源那么彻底的 https://woshipm.com/ai/6469597.html
• appinn Strata 现已支持 macOS，还能运行更轻量的 Qwen3.6-35B-A3B 大模型了 https://appinn.com/strata-macos-35b-model
• ifanr 英伟达RTX Spark笔记本首发实测，我实现了WorkBuddy额度自由 https://ifanr.com/1684682
• woshipm 实测Step 5 Preview，又一个能打的国产模型来了 https://woshipm.com/ai/6467696.html
• woshipm 一手实测 Step 5 Preview，国产 AI 模型卷麻了！ https://woshipm.com/ai/6467526.html
• appinn 8GB显存就够，Strata 让普通电脑也能运行千亿参数大模型 https://appinn.com/strata