Open Source Radar — October 23: MOSS-Image, MiniMax weights
Four releases in about 48 hours, three of them out of China, and the same sequence each time: the API ships first, the weights follow days later. That is the new default, and it makes the open-weight gap a question of latency rather than capability.
MOSS-Image (5B params, Apache 2.0 code) — OpenMOSS and Mosi released a 5B diffusion transformer trained from scratch, and QbitAI reports 1.3 million downloads in its first four days across the Hugging Face mirrors, plus the top of the site's text-to-image trending board. It is aimed at design work rather than pretty pictures: a frozen Qwen3.5-9B text encoder, prompts up to 4,096 tokens, and structured JSON layout input with bounding boxes, so a headline and a product land where you asked. The same QbitAI write-up puts it first among open-weight models on the LLTB text-rendering benchmark at 96.3% coverage, and second only to GPT Image 2 in a blind design eval at 1,698.70 Elo. The community is already outrunning the paper — an 8-step distilled variant has passed 75,000 downloads on Hugging Face, next to GGUF, INT8 and MLX builds.
MiniMax M3.1-Flash (428B MoE, weights published) — the model that went on the API on October 19 has its weights on Hugging Face as of this morning: 428 billion parameters in total, 23 billion active per token, native text, image and video input, and a 1M-token context. MiniMax's own card claims 73.7% on DeepSWE v1.1 — behind DeepSeek V4.1-Flash-1017 at 77.0, ahead of Kimi K3 and GLM-5.3 — with 61.7 on OSWorld 2.0 and 82.2 on CyberGym, all vendor-run numbers. The engineering worth reading is in serving rather than scores: sparse attention through every layer, quantization-aware training inside the routed experts, and a draft head for speculative decoding, which is why the checkpoint lands near 220 GB instead of something no one can host. The license is MiniMax's own, not Apache — read it before you build a product on it.
Beam 1.0 (501B MoE, Apache 2.0) — Reflection shipped a 501B-parameter mixture-of-experts that activates 23B per token, with a 1M-token context and genuine Apache 2.0 terms: distillation, fine-tuning and merging are explicitly permitted, which is not true of every "open" release this month. Weights come in BF16, FP8 and NVFP4 builds. The claim worth testing is token economy rather than benchmark placement — Reflection says Beam matches leading open models using roughly three times fewer reasoning tokens, and on long agent runs that is the number that shows up on the invoice. It is English-first, so it is a poor fit for multilingual products.
λ0 (openlambda) (Apache 2.0, code and data pipeline) — Alibaba's Amap published a whole-body humanoid foundation model trained from human movement instead of robot teleoperation, QbitAI reports: collectors wear consumer gear and walk through ordinary rooms, taking a demonstration from about 100 seconds down to 29. Amap claims state of the art on the SIMPLE simulation benchmark, a 22.5% lift in real-robot task success over the strongest baseline, and zero-shot transfer on seven loco-manipulation tasks. What is actually open is the training code, the capture-hardware rig and a sample of the 500-hour HumanVerse-500 dataset — not the weights, so the honest description is a replicable pipeline, not an open model. It is still the most useful thing published this week for a team collecting humanoid data.
If you had to bet on one of these four still being in use a year from now, which would it be? Write to us via our contact page.
