DeepSeek V4 Flash goes MIT: 13B active beats its own 1.6T flagship
Re-post-trained, not redesigned: same 284B/13B-active weights, now MIT-licensed and ungated, ~$0.14/$0.28 per M, and self-hostable on one 128GB box.
Copy markdown
Open weights, MIT, ungated
DeepSeek posted the full 284B-parameter (13B active) V4 Flash to Hugging Face under an MIT license with no gate — commercial use, fine-tuning and redistribution all allowed. Quantized GGUF builds are already up for llama.cpp.
13B active outscores the 1.6T Pro
On all nine agentic benchmarks DeepSeek published, the 0731 build beats its own V4-Pro flagship: Terminal-Bench 2.1 hits 82.7 (from 61.8) and DeepSWE jumps to 54.4 (from 7.3). The leap came from re-post-training, not a bigger model.
Agent loops without a GPU budget
The API runs ~$0.14 per M input / $0.28 output (cache hits $0.0028) — about a third of V4-Pro — so long agent loops are cheap without owning a GPU. A reasoning_effort flag toggles low/high/max to trade latency for depth.
Self-host 1M context on one box
Native FP4 experts plus FP8 layers fit the full 1M-token context in a single 128GB machine (~110GB at 3-bit), roughly 10% of V3.2's per-token compute. vLLM and SGLang enable speculative decoding with one flag.
The catch: it burns output tokens
Flash spends ~2.6x the median output tokens to finish a task, so per-task cost runs higher than the sticker price — and Claude Opus still tops all nine benchmarks. Read it as broadly competitive, not frontier-beating.
Elsewhere: Copilot CLI adds live tool timing
GitHub Copilot CLI v1.0.78 (Aug 3) now shows per-tool run timings live and auto-maintains first-party plugins, following recent grok-4.5 and Claude Opus 5 support in the CLI.