Skip to content
AI-04

Open-Source AI Models

Benchmarks, licences and deployment notes for open-weight models.

Recent

  1. Research noteMD-2026-0216

    Swap two cups and transformers lose the ball

    Ai2's StateBench finds open transformers fail at harder state tracking, while hybrid models trained on the tasks generalise much further.

    2 minSource: Hugging Face, Ai2

  2. News briefMD-2026-0215

    Built as a teacher model, it leads most of its class

    Nanbeige released a 33B model under Apache 2.0 as the teacher for smaller agents, with results against Qwen and Gemma models of its size.

    2 minSource: Hugging Face

  3. News briefMD-2026-0214

    816 billion parameters, 42 billion awake at once

    Meta's Muse Spark Open routes each token to 8 of 256 experts and takes text, images, audio and video as input.

    2 minSource: Hugging Face

  4. News briefMD-2026-0213

    A training stack rebuilt around sparse models

    Ai2's Olmo-core 3 keeps experts resident on GPUs instead of regathering weights, and reports 2.7 times the throughput of its predecessor.

    2 minSource: Hugging Face

  5. News briefMD-2026-0212

    A 280M drafter for a 3B vision model

    Liquid AI's draft model for LFM2.5-VL-3B decodes up to 3.13x faster on an M5 Max; end-to-end gains are smaller because prefill is untouched.

    2 minSource: Hugging Face

  6. News briefMD-2026-0211

    A quantisation format crosses libraries

    Transformers can now load GGUF checkpoints directly, reusing llama.cpp's own kernels rather than reimplementing them.

    2 minSource: Hugging Face

  7. News briefMD-2026-0210

    Granite 4.2 adds a switch for thinking

    IBM ships 3B, 8B and 30B under Apache 2.0, with chain-of-thought that can be turned off and a low-effort mode for simple queries.

    3 minSource: Hugging Face

  8. BenchmarkMD-2026-0209

    Three hundred million parameters doing the guessing

    Liquid AI's DSpark draft models are around 300M parameters and five decoder layers, and report up to 3.18× throughput on an H100 and 2.87× on a MacBook.

    3 minSource: Hugging Face

  9. BenchmarkMD-2026-0208

    Recovering most of what quantisation costs

    Liquid AI's Q4_0 checkpoints hold about 97% of full-precision accuracy while decoding 3–33% faster than the quantisations they replace.

    2 minSource: Hugging Face

  10. Field reportMD-2026-0207

    Qwen 3.8 27B thinks too hard by default

    Apache 2, 27 billion parameters, runs on a laptop — and spends 22,276 reasoning tokens on a request that needed almost none.

    2 minSource: Simon Willison

  11. News briefMD-2026-0206

    A 3B vision model that fits in 3GB

    Liquid AI's LFM2.5-VL-3B targets screen understanding on device, and publishes speed figures for CPUs rather than only for datacentre parts.

    2 minSource: Hugging Face

  12. AnalysisMD-2026-0205

    The licence is the specification

    Open weights are not open source, and the difference decides what you are allowed to ship.

    10 min

  13. BenchmarkMD-2026-0199

    Quantisation costs more on reasoning than on recall

    Four-bit weights barely dent factual tasks. Multi-step problems are a different story.

    7 min

  14. Field reportMD-2026-0193

    Checkpoint provenance is mostly unverified

    Teams pull weights from public hubs and rarely check what they received.

    6 min

  15. AnalysisMD-2026-0188

    Fine-tuning is back for narrow tasks

    Prompting won the general case. For classification and extraction the economics reversed again.

    8 min

  16. Research noteMD-2026-0182

    Model cards have stopped being useful

    The format survived. The candour did not.

    5 min