OSS

DeepSeek dropped V4 in two preview models today — V4-Pro (1.6T total parameters, 49B active) and V4-Flash (284B, 13B active), both 1M-context MoE under an MIT license. Pro is now the largest open-weights model in circulation, larger than Kimi K2.6 and GLM-5.1, and its $1.74-per-million-input pricing undercuts every Western frontier. Self-reported benchmarks trail GPT-5.4 and Gemini-3.1-Pro by three-to-six months — but at this price gap, the spread barely matters for most production workloads.

The clearest practitioner teardown of DeepSeek V4-Pro and V4-Flash that exists the morning they shipped — pricing comparison table, quantization notes, MoE activation math, and the pelican-on-a-bicycle test. The single pricing table alone reframes every model-selection conversation you'll have next week: Flash at $0.14 input beats GPT-5.4 Nano, Pro at $1.74 input beats every other frontier model outright.

Arcee released an open-source reasoning model optimized for multi-turn agent workflows, tool usage, and instruction-following without sacrificing cost efficiency.

George London argues that AI agents can revitalize the free software ecosystem by reducing the friction that has kept open source from reaching its full potential.

Mistral released a 4B-parameter multilingual TTS model supporting nine languages, optimized for on-device deployment with minimal audio samples for voice cloning.

Analysis of how the monetizable capability gap between frontier labs and open-source models is narrowing faster than raw capabilities, with implications for pricing and moat strategy.

Open-source models are rapidly closing the gap with frontier alternatives, and this indie developer makes the case that local processing is becoming viable for serious work.

NVIDIA released Nemotron 3 Super at GTC 2026, a hybrid Mamba-Transformer MoE model with 120B parameters that activates only 12B at inference time. It ships with a 1M-token context window and is optimized for multi-agent agentic reasoning tasks. Open weights are available on Hugging Face.

Karpathy's new framework automates the grind of model training experiments — hyperparameter sweeps, ablation studies, and result analysis. Open source and immediately useful.

Alibaba's compact 9B-parameter open-source model beat OpenAI's much larger offering on key benchmarks, released under a permissive Apache 2.0 license.