Open Source Factory

MODELS & COSTS DeepSeek Flash Vision: API Access and Weights Are Separate Open Source Factory · Aug 31, 2026 · 5 min read An experimental API and a downloadable checkpoint create different ways to evaluate Flash Vision. The August 21 and August 31 dates refer to those separate releases. The model's text and multimodal scores remain DeepSeek's measurements until a matching independent run is available. In this article API and weight-release dates Text-agent results in the vendor table Multimodal-agent results Position within the V4 family API integration notes Architecture and deployment notes The gist: Same 284B / 13B-active Flash architecture with vision modules added and continued training. On text-only agents it stays on par with Flash 0731 (Terminal Bench 83.9 vs 82.7, DeepSWE 59.3 vs 54.4) — and on multimodal agents it jumps +10.3 on ApexBench to 36.5 , closing to within 2.9 of Opus 4.8. 1M context, MIT, 384 tokens per image at Flash pricing. The repository is d...

HARDWARE & DESIGN Local AI Hardware: Capacity, Bandwidth and the Cost of an Upgrade Open Source Factory · Aug 31, 2026 · 4 min read Before buying more memory, choose the model and job that will use it. A small fast text model, a large MoE and a video pipeline put pressure on different parts of a computer. A complete quote and a matched workload test are more useful than a ranking built from peak specifications. In this article Define the workload before comparing computers Keep the memory resources separate Ask for a matched performance test Compare the configuration you would actually buy Calculate ownership cost against the same work Specifications and related workload guides Correction, October 17, 2026: an earlier version ranked computers using forecast token speeds for an unconfirmed Qwen3.8-35B-A3B checkpoint and simplified large-model fit assumptions. Those rankings had already been withdrawn in an October 3 note. The unsupported tables and five derived charts have now been...

HARDWARE & DESIGN 35B MoE Memory Planning: What 3B Active Does Not Tell You Open Source Factory · Aug 31, 2026 · 4 min read Updated October 3, 2026 · Author calculations and a planning guide · No hardware benchmark Correction to the earlier version: this URL previously discussed a projected Qwen3.8-35B-A3B release and estimated Mac mini speeds. I could not establish those model-specific speed, context-memory and purchase claims from an official released checkpoint and reproducible measurements. They have been withdrawn. This article now explains the calculation you can check yourself; it does not announce that model or recommend a Mac configuration on the basis of a forecast. Three billion active parameters do not turn a 35B MoE into a 3B download. The inactive experts still need storage, and the application needs memory beyond the weights. Work through that budget before using an active-parameter label to select a 24GB computer. In this article Calculate the raw weights befor...

MODELS & COSTS Modified GLM-5.3-Flash: Refusal Scores and Their Limits Open Source Factory · Aug 30, 2026 · 3 min read A lower refusal rate describes behavior on a particular test set. It does not measure every capability or make a modified checkpoint interchangeable with the original. The GLM modifications below should be compared through the teams' stated methods, serving conditions and limits on independent evidence. In this article Refusal changes reported in the weights Reported capability and serving costs How this release fits earlier modifications Linked source documents The numbers OrcaRouter published are the hook. Refusal behavior collapses from 96%, 93%, 97%, and 93% down to 11%, 12%, 15%, and 18% across the four categories they measured, and XSTest — the rate of wrongly refusing harmless requests — drops from 2.4% to 0.4%. Their pitch: a stock GLM-5.3-Flash was over-refusing a lot of benign-but-flagged requests, copyright questions being the clearest real case. ...

IMAGE, VIDEO & AUDIO H3 Max: The Hosted Retrain and the Open Base Open Source Factory · Aug 28, 2026 · 5 min read fal's H3 Max is a post-trained H3 service, not the same deployment as the downloadable base model. Its reported speed and arena placement make an API comparison worth considering. Local editing, released artifacts and hosting requirements need a separate check. In this article The base model and fal's post-training Output examples Latency claims and serving conditions The dated arena rankings Price and launch conditions Generation, editing and local use H3 Max uses the 33B H3 base, with post-training aimed at prompt adherence and visual quality. At the article's leaderboard snapshot, Artificial Analysis placed it first in image-to-video and third in text-to-video with audio. Its 768p limit and unreleased weights distinguish it from the base H3's local and editing workflows. From open weights to a #1-ranked retrain in about a month (author's mile...

HARDWARE & DESIGN Nvidia–Hugging Face: What the Reported $12.9B Deal Would Change Open Source Factory · Aug 27, 2026 · 4 min read October 3, 2026 update: This article describes the August report. Subsequent reporting says NVIDIA confirmed the agreement; that does not mean the acquisition has closed. The original uncertainty wording below belongs to the publication date. Image: Hugging Face logo (huggingface.co). Reports of Nvidia buying Hugging Face matter because the hub distributes models used far beyond Nvidia hardware. The accounts disagreed on price and whether anything was signed. This article examines what such a deal could change, while keeping the reported talks separate from a completed acquisition. In this article The strategic case described in the reports The cloud business alongside the model hub How the reported prices compare Questions a completed deal would raise Hugging Face's reported scale Linked source documents The reports do not establish a comple...