This observation reviews public technical material around the period from 7:36 am on 24 October to 7:36 am on 25 October 2026 in Australia/Sydney and includes two model updates whose delivery status can now be verified. Both complete follow-up steps recorded earlier in this series: one moved from hosted access to downloadable weights, while the other moved from an announced delivery date to general availability.
MiniMax releases M3.1 Flash weights and local deployment material
MiniMax first made M3.1 Flash available through its API, M Plan and MiniMax Code on 19 October, when the public status still said that weights would follow. On 23 October, MiniMaxAI's official Hugging Face repository made the model files, DSpark draft head and local deployment instructions available. The documented serving paths include SGLang, vLLM, TokenSpeed and multi-chip deployment through FlagOS. The model card lists about 428 billion total parameters, around 23 billion activated per token, text, image and video input, and a context window of up to one million tokens. Usable context and throughput depend on GPU memory and serving configuration.
The weights are governed by the MiniMax Community License rather than an unconditional standard open-source licence. Non-commercial use is free. Commercial products must display “Built with MiniMax M3.1-Flash”; products earning no more than US$20 million a year require a one-time notice to MiniMax, while those above that threshold require prior written authorisation. The licence also excludes military use and other listed purposes. Capability scores in the model card remain developer-reported. TechNode independently confirmed that the weights and deployment material were public on 23 October, but this review found no independent reproduction covering all capability claims.
NVIDIA delivers Nemotron 3.5 Bolt E4B 4GB as scheduled
When NVIDIA released Nemotron 3.5 Super VL and Bolt 74B-A9B on 15 October, it listed Bolt E4B 4GB for delivery on 23 October. On that date, NVIDIA's model cards marked Nemotron 3.5 Bolt E4B 4GB as generally available and provided a Q4_K_M quantised file, a full-precision BF16 version, a DFlash draft model and local-run instructions. The new fact is that the previously scheduled small model is now downloadable; it is distinct from the 15 October release already recorded by this series.
The model is post-trained from Gemma-4-E4B-it and has eight billion total parameters with four billion effective parameters. The 4GB release uses Q4_K_M precision, permits up to 64K tokens for input and output, supports switchable reasoning and multi-turn tool use, and targets local voice and device assistants, interactive personas and embedded interfaces. It uses the OpenMDW 1.1 licence together with Apache 2.0 terms inherited from Gemma 4. NVIDIA's benchmark figures come from its own evaluation harness and do not yet have complete independent reproduction. Unsloth independently documents the downloadable files and llama.cpp path, without validating all vendor performance claims.
Sources
MiniMaxAI, MiniMax-M3.1-Flash model card, 23 October 2026: https://huggingface.co/MiniMaxAI/MiniMax-M3.1-Flash
MiniMaxAI, MiniMax Community License, repository version released 23 October 2026: https://huggingface.co/MiniMaxAI/MiniMax-M3.1-Flash/blob/main/LICENSE
TechNode, MiniMax releases open weights for 428-billion-parameter M3.1 Flash model, 23 October 2026: https://technode.com/2026/10/23/minimax-releases-open-weights-for-428-billion-parameter-m3-1-flash-model/
NVIDIA, Nemotron-3.5-Bolt-E4B-BF16 Model Card, 23 October 2026: https://build.nvidia.com/nvidia/nemotron-3.5-bolt-e4b/modelcard
NVIDIA, NVIDIA-Nemotron-3.5-Bolt-E4B-4GB model card, 23 October 2026: https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Bolt-E4B-4GB
Unsloth, NVIDIA Nemotron 3.5 Bolt: Run Locally Guide, version reviewed 24 October 2026: https://unsloth.ai/docs/models/nemotron-3.5-bolt
Discover more from Geoffrey Chen
Subscribe to get the latest posts sent to your email.