Added, by axis
Axes with the most records come first. A record that spans several axes is counted once, under its main axis.
⑦3 records
Multimodal
- Oct 15, 2026Black Forest Labs — Black Forest Labs released the open weights of FLUX.3 Image Klein — one model handles making a picture from text, editing with several reference pictures, and layout control that places regions by writing coordinates into the prompt SourceIt shipped two versions, a 16-step distilled one and a 50-step base one, together with FP8 and NVFP4 quantized weights and a DreamBooth LoRA training script. Only a speed figure was given — the distilled version makes a 1024×1024 picture in about 2 seconds on a single H100 — and no quality scores
- Oct 15, 2026Luozheng Qin et al. — MJev-Spatial released — instead of writing out an answer to a spatial question as text, it solves it as a multiple-choice problem, scoring options, numbers and 2D coordinates with projections the model already has and settling the answer in a single computation SourceThe authors reported that, trained on 12.3 million question-answer pairs on top of Qwen3.5-4B, it ranked No. 1 among the 8 models compared, with an overall 65.64%, and was No. 1 on 7 of 12 benchmarks. They reported it was 2.5 times faster than the fastest generative approach and 5.6 times faster than the model it was built on
- Oct 14, 2026xAI — xAI released Grok Voice 3.0 — a model that answers speech to speech directly without turning speech into text, offered in the Voice Agent API at $0.08 per minute, with existing grok-voice-latest usage switched over automatically SourceCompany-reported figures: an Artificial Analysis speech-to-speech index of 86.9 (GPT-Live-1 83.2, Gemini 3.8 Live 82.6), Full Duplex Bench 97%, and response latency under 1 second. xAI said it is used in Tesla vehicles and for Starlink customer support, where about 70% of roughly 20,000 calls a day are finished without a person
⑨2 records
Open-Weight Lineage
- Oct 16, 2026Reflection — Reflection launched Beam, an open-weight model — a sparse mixture-of-experts architecture with 501 billion total and 23 billion active parameters per token, released under Apache 2.0 with a context of 1 million tokens SourceIt is the first model from a new US lab. Company-reported scores were SWEBench Verified 81.6%, GPQA Diamond 90.6%, AIME 2026 97.7% and Terminal Bench v2.1 80.1% (on BF16), and it ships in three versions, BF16, FP8 and NVFP4, with the NVFP4 checkpoint at about 280GB
- Oct 15, 2026Mistral AI — Mistral AI released Timestral, a time-series forecasting model — four sizes, 5M, 28M, 123M and 209M, under a modified MIT license, producing probabilistic forecasts over 21 quantiles from a context of 32,768 time steps on CPU alone without a GPU SourceCompany-reported, the 209M model ranks No. 1 in the pretrained zero-shot categories of GIFT-Eval, TIME and fev-bench, and the 28M model beat Chronos-2, which is about four times larger. On an 8-thread server CPU the 5M model takes about 10ms and the 209M model about 140ms. A technical report was only announced
⑫2 records
Training Data & Post-Training
- Oct 16, 2026NVIDIA, AI4Bharat (IIT Madras) — NVIDIA announced the coming release of a 2.8 trillion token pretraining corpus for 22 Indian languages — Nemotron-pretraining-Indic-v1, built with AI4Bharat at IIT Madras, comes together with an open generation pipeline that combines translation, quality scoring and restoration SourceCompany-reported, it covers 34 language and script combinations with 2.601 trillion tokens in native scripts plus 224.6 billion in Latin script, and includes 438.4 billion restored tokens kept only where the quality score was 75 or above. Continuing the training of Nemotron 3.5 Super VL on part of this corpus raised five Indic benchmarks by 4.98 to 9.82 percentage points, with the caveat that this is not the effect of the synthetic corpus alone
- Oct 15, 2026Bingxi Hou et al. — Waypoint-OPD released — rather than taking the teacher model's change in probabilities before and after reinforcement learning as it stands, it compares them with a midpoint of the same training run and, where the direction of change before and after disagrees, damps that signal as it transfers it to the student model SourceIt starts from the point that looking only at the endpoints lets preference changes from early and late training cancel each other out and hide the parts where the teacher changed consistently. The authors reported that with a Qwen3-4B teacher and three Qwen3-family students, in all six settings compared across the two domains of math and code, it had the highest domain average among the methods compared
⑯2 records
Evaluation & Benchmarks
- Oct 15, 2026Dongxu Yang et al. — FDE-Gym released — a benchmark modeled on the work of an engineer posted to a customer, with 95 delivery tasks drawn from 20 situations across 18 industries that have to be solved while asking after vague requirements and constraints nobody stated SourceThe authors reported that in the setting where the constraints have to be found by the model, even the best configuration averaged only 0.56 points and solved just 14 of the 95 tasks. Even in the 23 tasks where every necessary fact was obtained from the stakeholder, only 10 scored 0.8 or above, and they concluded that finding the unstated constraints is the bottleneck
- Oct 13, 2026Xuyang Liu et al. — OmniCompass released — a benchmark that measures video understanding by splitting it into three parts, when something appears, where it is and what is happening, with a separate track that uses sight only and a track that also uses sound SourceThe authors said the vision track was built from 820 videos and 2,460 questions and the track that also uses sound from 240 videos and 720 questions, all annotated by people and checked several times, and that 50 models were evaluated. They reported that open models fell behind commercial models especially on tasks that use sound and screen together, and that answers often differed between pointing to the same object in text only and giving a picture along with it
③1 record
Tools & Agents
- Oct 15, 2026Hanzhang Zhou et al. — OpenGCUA released — a study that leaves every component of a general-purpose computer-using agent open, combining screen control with code execution and moving the job of judging what to look at and how far the work has come from harness rules to the model SourceThe authors reported that on OSWorld-v2 the average partial score of the 8 models evaluated rose by 19.6 points and that steps per task fell by 53.0% on average compared with running screen control alone. They released the implementation, prompts and evaluation logs together with an early version of a harness for real devices
⑤1 record
Loop & Harness Engineering
- Oct 14, 2026Cohere — Cohere published experiments that group coding agents into teams — with a minimal harness in which a lead agent hands out the work, it varied team size, model mix and whether the work ran asynchronously, and had 200 ProgramBench tasks rebuilt from scratch SourceCompany-reported, a single Sol agent at 52.1% rose to 61.8% once it worked as a team, and a team of a Sol lead with 16 Luna planners reached 61.8% for $155, about 30% of the cost of an all-Sol team that delegated every stage ($510, 63.2%). Cohere added the caveat that teams cost 3 to 7 times more than a single agent and that the score on one task swings by about 6 points from run to run
Showing 11 here; there are 6 more, including 6 other axes. Open the Advanced page to see them all.
How this relates to the main moduleThis page holds
only what was added that week. For the whole story, open
the main module — it is
one latest edition, revised every week, while weekly editions
stay as they were.