Tag
#Open source
From News
News · 2026-11-20
Runtime-specific LoRA raised coding scores while cutting agent work by half
Hugging Face trained separate Qwen3.8-27B LoRA adapters inside Pi, Hermes Agent and OpenCode. In their target runtimes they improved two coding benchmarks and used 53 to 72% fewer tool calls on jointly solved tasks.
Read →News · 2026-11-18
Transformers runs a 180B MoE on a 32GB Mac, trading memory pressure for SSD traffic
A Hugging Face experiment ran the 180 billion parameter Qwen3.8-Flash-Next on an M5 Mac with 32 GB of unified memory. It streams selected experts from SSD and reached 13.9 tokens per second on short generation, about 6% behind the specialised Strata engine.
Read →News · 2026-11-18
Rumble-2-Flash packs transcription for 45 languages into 0.5B parameters
Reson8 has released the Rumble-2-Flash weights under Apache 2.0. The 0.5B parameter model averages 6.60% WER on the English Open ASR Leaderboard and targets inexpensive local transcription without a custom runtime.
Read →News · 2026-11-18
IBM puts time-series forecasting directly into streaming SQL
IBM Granite time-series models now run as built-in Apache Flink functions on Confluent, in the cloud and on premises without GPUs. The practical value is not another foundation model, but a forecast produced inside the same data stream that can immediately trigger an action.
Read →News · 2026-11-12
LoRA fixed shoe soles across 188 objects while leaving textures alone
Hugging Face has released containers and scripts for fine-tuning TRELLIS.2 on custom 3D assets. An experiment with 188 footwear models raised F-score from 0.276 to 0.803, but the broader lesson is restraint: train only the components for which the dataset provides the right signal.
Read →News · 2026-11-11
Pooling-aware training retains 97.8% of retrieval quality with one quarter of the vectors
Hugging Face has demonstrated a training method that teaches a multi-vector embedding model to survive the merging of similar token vectors. With 4x compression, the experimental model retained 97.8% of its original nDCG@10, compared with 94.6% for the ordinarily trained variant.
Read →News · 2026-11-05
NVIDIA opened four world simulators. The entry ticket is a 16B model and formidable hardware
NVIDIA released four open-weight Cosmos 3 Nano checkpoints for robotics and autonomous driving. They turn actions and spatial controls into continuing video or LiDAR, but the published environment was tested on one GB200 and NVIDIA itself warns about physical errors.
Read →News · 2026-11-06
ToyBlocks gives robot learning a controlled laboratory, with a 5.5-point sim-to-real gap
Ai2 released the open ToyBlocks simulation with 85,336 demonstrations and nearly 1,600 hours of robot data. Its hierarchical system scored 55.5% in simulation and 50% on a physical robot, although only in the easiest scene configuration.
Read →News · 2026-11-05
Phonon fits local TTS into 65 MB, but its licenses split the model family
Phonon offers local TTS models of roughly 65 MB and 75 MB across five languages and more than 10 compute backends. The English 31m model permits commercial use under Apache 2.0, while multilingual options carry the more restrictive CC BY-NC license.
Read →News · 2026-11-03
IBM taught a 20B model to build decks. The verification loop does the decisive work
IBM Research fine-tuned the 20B open-weight Palette model for decks in the IBM Design Language. It ranked second on average in PresentBench, but the stronger result is a pipeline that separates planning from drawing and accepts repairs only after measurable improvement.
Read →News · 2026-11-04
Open models are growing faster than the market can measure them
Interconnects estimates the open model inference market from public figures by Fireworks, Baseten, Together AI and OpenRouter. Fireworks reports more than 40 trillion daily tokens for customized models alone, but missing cloud breakdowns prevent a clean comparison with closed models.
Read →News · 2026-11-03
StateBench shows that a long context window is not reliable memory
Ai2's StateBench measures whether a model can preserve variable state across long sequences of changes. Four frontier models fell to 13% accuracy or lower on the longest, hardest tasks.
Read →News · 2026-10-30
CleanYourTV gives Android TV owners a way to reclaim the home screen
The open-source project disables selected ad and tracking services on Android TV and Google TV, then replaces the crowded launcher with a simpler one. Its rapid growth reflects a familiar frustration: buying a TV should not mean accepting ads in its menus.
Read →News · 2026-10-29
IBM Patches Granite Switch to 4.1.1 by Swapping Adapters Without Retouching the Base Model
IBM released Granite Switch 4.1.1 as a backward-compatible patch that updates its 10-language speech recognition model and decision adapter while leaving the base model untouched. The release demonstrates that modular adapter architectures allow teams to ship incremental capability upgrades on a single endpoint without recomputing core weights or forcing user migrations.
Read →News · 2026-10-29
Parakeet transcribes an hour of audio on a Mac in 20 to 40 seconds. Hugging Face slipped it under Apple's own API
Hugging Face released two Swift packages. AnySpeechRecognizer keeps Apple's speech-to-text interface but runs Parakeet or Qwen3-ASR locally instead of Apple's engine, and adds speaker diarization. AnyDecisionModel answers yes-or-no style questions with probabilities instead of generated text.
Read →News · 2026-10-27
Change a few words in the prompt and an AI novelty judge drops from 93.5% to worse than a coin flip
Ai2 measured how reliably language models can tell whether a research idea is new. After a small prompt change GPT-5.4 fell from 93.5% to 40.9% accuracy, and a plain Claude Opus 4.6 prompt costing $2.59 beat two expensive purpose-built evaluators.
Read →News · 2026-10-27
Liquid AI squeezed six models into 2 billion parameters, but big companies may not deploy them commercially
Liquid AI's LFM3 family puts text, vision, audio and decision models on one backbone, and the text model fits in about 1 GB at 4-bit precision. The weights are free to download, but the license allows commercial use only for companies with annual revenue under $10 million.
Read →News · 2026-10-22
IBM folds four Granite models into one forecast. It beats the best of them by two thousandths, and the real win is not having to choose
IBM Research has released Granite Timeseries Ensemble, a recipe that combines four pre-trained time-series forecasting models without any extra training. On the GIFT-Eval benchmark it beat each of its members on point accuracy, but the best one by only two thousandths. The real value is that a data team no longer has to guess in advance which model suits its data.
Read →News · 2026-10-21
Photoshop in Rust in two weeks. ArtCraft has thirteen apps and rates itself at 45 %
Brandon Thomas set Claude Opus 5.5 to rewrite Photoshop, Premiere, Lightroom and other programs in Rust. Three weeks later there are thirteen open-source apps and more than 70,000 stars for PhotoCraft alone. The project's own roadmap, however, says hundreds to thousands of agent-hours stand between it and real work, and lawyers disagree on whether this is a clean room.
Read →News · 2026-10-22
REA collected 300,000 stars in two weeks. Its most valuable feature is a list of what it doesn't know
The open-source tool Reverse Engineer Anything connects coding agents to Ghidra, Hopper and IDA through a single MCP server. On GitHub it grew from 41 stars to more than 360,000. An independent test, however, found that on small programs an agent manages just as well without it, and a fake copy carrying a trojan rode along on its popularity.
Read →News · 2026-10-20
DeepSeek opened up a cheap model and its inference stack. US federal agencies are now barred from buying it
In its 25th open artifacts roundup, Interconnects describes a week in which DeepSeek released V4.1 Flash under the MIT license and open-sourced its DeepInfer inference stack. At the same time, the US Department of Homeland Security shut it out of federal procurement and China proposed export approvals for frontier models. Open weights are entering an era where provenance decides how they can be used.
Read →News · 2026-10-20
FairStress: LLMs abandon correct decisions more readily depending on who benefits
Ai2 released FairStress, a benchmark of more than 13 million tests that measures how easily a language model drops a well-supported decision about people. Models including Claude Opus 5 and GPT-5.5 gave way unevenly: arguments favoring a minority-coded candidate swayed them 8.8 to 42.4 percentage points more often.
Read →News · 2026-10-20
IBM gave GPT-4.1 a cheat sheet built from its own mistakes, and the agent beat frontier models on τ²-bench retail
IBM Research had the older GPT-4.1 model write rules from its own attempts and added them to the prompt. The customer service agent reached 67.5% Pass^4 on τ²-bench retail, ahead of every leaderboard entry, including a fine-tuned model.
Read →News · 2026-10-07
Nemotron 3 for IOI and IMO: The end of general models, specialization through SFT and RL decides
Nvidia's Nemotron 3 has reached gold-medal level at both the IOI 2026 and IMO 2026 competitions. Instead of building a new massive foundation model, it demonstrated a reproducible fine-tuning recipe (SFT and RL) on top of an existing base, complemented by a feedback-driven inference loop.
Read →News · 2026-10-07
Open d1 makes edge decisions in 16 ms, but multimodal evidence is still missing
Liquid AI has released the open-weight d1-3B and experimental d1-omni-600M for structured decisions without token generation. d1-3B answers in 16 ms on a Jetson AGX Thor, but this release includes no vision or audio benchmarks.
Read →News · 2026-10-03
ThinkingBox grades agents by the database, not by confident answers
Microsoft and Hugging Face have made ThinkingBox available, a benchmark of 507 stateful business tasks that runs each task 20 times and inspects the actual backend outcome. It shows why a successful tool call or polished answer does not mean the job was completed.
Read →News · 2026-10-02
AI writes 60% of Airbnb code, but handoffs are the bigger change
Airbnb CTO Ahmad Al-Dahle says AI authors 60% of company code and average pull request throughput per engineer is up 1.6 times. The deeper change is a move from documents and handoffs to prototypes, the Everest context layer and use-case-specific evals.
Read →News · 2026-10-02
AstaBrief writes cited research reports in 51 seconds with 8 billion parameters
Ai2 has released the weights for AstaBrief 8B, which generates cited scientific reports in Fast mode in an average of 51.1 seconds versus 178.5 seconds for Thinking mode. Its speed and self-hosting option are compelling, but much of the evaluation dates from 2025 and the small human study covered only 14 questions.
Read →News · 2026-09-28
Holo4 spans screens, code and APIs, but its benchmarks run on different tracks
H Company released the Holo4 27B and 35B-A3B models, combining GUI control, code execution and MCP or API calls in one agent. Open weights and published trajectories improve auditability, while results from different harnesses and task sets still complicate direct comparisons with closed models.
Read →News · 2026-07-12
White House tests the waters for banning frontier open-source models
The US government is considering how to hit the brakes on open AI models before they catch up to the closed frontier. The official narrative is about security risks from China, but the reality will hit everyone.
Read →From the Library
Library
Honcho: memory that remembers people, not just text
Honcho is an open-source memory layer for AI agents from Plastic Labs. It stores conversations as messages, derives conclusions about people and agents in the background, and returns model-ready context on request. It fits agents that must pick up where they left off across days, tools, and interfaces.
Read →Library
Koog and Kotlin AI agents — what it is and what it is for
Koog is JetBrains’ framework for building AI agents in Kotlin and Java. It focuses on practical architecture: strategies, tools, memory, tracing, long context and JVM production integration.
Read →