Neurodigest for Two Weeks (#128)

Neurodigest for Two Weeks (#128)

LLM

• GLM 5.3 Flash — the former Ox Alpha has become an open multimodal model 320B-A18B. The API costs $0.15/$0.5 per million tokens.
• Qwen 3.8 Flash Next — 125B-A6B plus 51B N-gram parameters on Qwen 4 architecture. Weights already support vLLM and SGLang.
• GigaEmbeddings — three open MIT models from 480M to 10B for RAG and semantic search.
• GPT 5.6 Sol has decreased in price — for three months, the price has been reduced to $4/$20 per million tokens with a context of up to 272K.

Generative Models

• MiniMax H3 Max by fal — this version generates five-second videos in about three seconds and leads the ranking of Artificial Analysis for image-to-video.
• FLUX Video Upscale — a video model that natively upscales videos to 4K in either precise or more creative mode.
• MAI-Image 2.6 Test — four runs produced almost identical characters, but half of the frames had issues with fingers.
• MAI-Image 2.5 Pro — Microsoft has emerged as a leader in the t2i arena, but generating 1024×1024 costs about $0.11.
• MAGI-2 Preview — an open MoE video model 114B-A6B generates ten-second videos immediately with sound.
• Gemini Omni 1.1 Flash — the release version of the new omni-video generator from Google. It generates up to 10 seconds with an extension to 40 seconds while maintaining the context of the last 10 seconds.

Hardware

• Jalapeño Chip — OpenAI has shown the first measurements of its inference chip: higher performance per watt and lower latency compared to systems from InferenceX comparison.
• Mac Studio with M5 Ultra — up to 512 GB of unified memory with a bandwidth of 1.2 TB/s for local deployment of large models.
• Robotic Duck for $399 — Hugging Face and Pollen Robotics have opened pre-orders for a 25-centimeter bipedal Microduck.
• Cerebras CS-4 — a server on three WSE-3 Turbo has become twice as fast as CS-3 and is designed for scaling to models with 10T parameters.
• Unlocked CMP 170HX — a patch restores computational power to mining cards and opens up 40 or 64 GB of HBM2e.

Other

• Keenable Agent Search — the startup has emerged from stealth and claims a better price-to-quality ratio than Exa.
• Skill for ARC-AGI-3 — the executable reasoning check raised the Opus 5 result from 30% to 100% and questioned the usefulness of the benchmark.
• Hugging Face is looking for a buyer — according to Business Insider, the company evaluated interest in selling for about $13 billion.
• Nvidia and Hugging Face Deal — The Information reported an agreement for $12.9 billion, but there is no public confirmation from the companies yet.
• Nvidia and Poolside — Nvidia will pay $6 billion for a license for technologies and offered to transition more than a hundred employees from the startup.