Qwen3.8-Flash-Next now runs local: GGUF quants, 75GB floor
The 125B/6B-active Qwen4-preview MoE fits a 96GB workstation for coding and agents — plus Cohere's $1.50 doc parser and Gemini's new speech-to-text.
Copy markdown
The quant that fits a workstation
Unsloth's dynamic GGUFs run Qwen3.8-Flash-Next on as little as ~75GB of RAM or unified memory: the 72.5GB UD-IQ1_S holds ~77% of full accuracy, while the 111GB UD-Q4_K_XL reaches ~92%. A 96GB Mac or workstation is the practical floor. Launch it with llama.cpp or `unsloth run`.
What you're actually running
It's a 125B-total / 6B-active hybrid MoE (512 experts, Gated DeltaNet plus sparse attention) — a preview of Qwen4's architecture with a 262K-token context, under the permissive Qwen Community License. It's tuned for coding and agents: 62.5% on SWE-bench Pro, 91.9% on LiveCodeBench, and 73.5% on Toolathlon tool use.
Cohere Parse 5 undercuts Mistral OCR at $1.50/1K pages
A day after Mistral OCR 4.1 hit GA at $4 per 1,000 pages, Cohere shipped Parse 5 — a 2.3B vision model that turns PDFs and images into structured Markdown with bounding boxes at $1.50 per 1,000 pages. It's API-only (no open weights), also deployable on Bedrock, SageMaker and Azure Foundry, and trades a few benchmark points for the best price-per-page on RAG ingestion.
Speech-to-text gets shaken up
Google moved Gemini 3.5 Transcribe and Transcribe Live to GA — 85+ languages, speaker diarization, word-level timestamps and streaming over WebSockets. Meanwhile OpenAI set whisper-1, gpt-4o-transcribe and gpt-4o-mini-transcribe to retire on Feb 26, 2027; migrate to gpt-transcribe or gpt-live-transcribe. If you run voice pipelines, plan the swap now.