Search

Last 30 days

  1. Friday, November 20

    Superfluid by basecompute

    The local server now distributes inference across Apple, AMD and NVIDIA machines: six machines raise throughput from 112 to 534 output tokens per second for 64 simulated users, and shutting down a node causes no requests to be lost,…

    WavLate

    This open late-interaction model searches for sounds using text and for text using sound, with one vector every 0,31 seconds of audio and one per word; its author reports an average Recall@1 of 37,9 on AudioCaps and Clotho, the best of the…

  2. Thursday, November 19

    Atto

    NeuML releases a family of tiny document retrieval models, ranging from 840 parameters to 28 million: the 90,000-parameter dense model achieves an nDCG@10 of 0.4257 on a subset of BEIR, compared with 0.4802 for all-MiniLM-L6-v2, which is…

    ja-NCM

    A developer releases a small Japanese open-weight model that generates text through concepts rather than token by token: according to the developer, it is around 6 times faster than an autoregressive model and 2.4 times faster than a…

  3. Wednesday, November 18

    ANEMLL

    A research version splits Qwen3.8-27B between the Neural Engine and GPU of a Mac M6: a 64K prompt takes about two and a half minutes instead of five minutes on the Neural Engine alone; it requires disabling SIP.

    Samdei Type

    Two small models under Apache 2.0 (7.1M and 18.2M parameters) convert Khmer typed in Latin letters into Khmer script, with a character error rate of 0.052 compared with 0.066 for the best existing open model, according to the author.

    ViKernel Scanner

    VIDRAFT releases an open-source scanner that gives models mock tools: all 20 open models tested receive an F grade, with the median model attempting a dangerous action in 40 out of 100 tests.

    Uncensored Kimi K3.1 Pro

    Provider OpenUncensored serves an unrestricted version that handles 97% of the legitimate but borderline requests in its test set, compared with 83% for the original, with an acknowledged weakness in agent loops.

    Also about: Moonshot AI

    RoboTwin Unified

    Harpreet Sahota releases a FiftyOne import of 27,500 episodes and more than 6 million images of dual-arm manipulation across 50 tasks.

    Recapture

    Nutrient releases an open-source tool that turns an annotated page into realistic phone photos or scans while preserving the annotations, to train document-reading models.

    Local Kafka assistant

    In the third part of his series, Roman Melnyk replaces fine-tuning with document retrieval using unmodified Qwen3.8-27B: +1.48 across 40 questions, compared with +0.08 and -0.18 for his fine-tuned adapters.

  4. Tuesday, November 17

    OpenRouter's Auto Router

    According to the AutoBench benchmark, whose author also offers a router, the new Auto Router chooses the right model in 65.2% of cases on the high setting, compared with 79.1% for the old one; its other settings remain at chance level.

    Also about: OpenRouter

    Uncensored models as agents

    OpenUncensored, which sells access to these models, tested 32 of them on five tasks involving tools: those derived from the best original models complete 38% of tasks, the others 28%.

    TuHaMa

    Turkish NLP Suite publishes 52,000 examples of Turkish hate speech targeting 12 groups, drawn from the Ekşi Sözlük forum, presented as a test of models' cultural understanding.

    AI Model Hacking Hackathon

    VIDRAFT opens a browser-based ViKernel red teaming competition, spanning four seasons, with a $4,500 prize pool.

    LeWAM

    Kevork Sulahian asked GPT-6 Astra to fix the LeWAM robotic policy by directly editing its weights: across 300 previously unseen starting states, none of the seven edits outperforms the original policy (-2.3 to +3.0 points).

    TartanGround for FiftyOne

    Voxel51 converts six ground robot trajectories from CMU AirLab's dataset (hospital, port, marsh, Japanese city, factory, downtown), with camera, segmentation, lidar, pose and IMU data.

  5. Monday, November 16

    projectsim (Kaedim)

    More than 200 3D objects ready for robotics simulation, including 140 articulated objects, are becoming free and are also joining NVIDIA's SimReady Central, where seven are already online.

    llama.cpp CI

    Faced with 3,181 pull requests in three months and the limits of GitHub's free plan (10 GB of cache, 20 concurrent jobs), the team now runs only jobs affected by the diff, stores its ccache caches in Hugging Face Buckets and runs most…

    Tiny Aya Global

    Five community posts, most from the Lexsi Labs hackathon, tackle Cohere Labs' small multilingual model: a database agent goes from a 0.5% to a 49.1% success rate over three rounds of self-repair, and refusal steering sharply reduces…

    Modular Diffusers

    A guide shows how to use a model that has not yet been integrated into Diffusers by converting its weights and publishing blocks on the Hub; the author assigned the work to Claude Code with Claude Opus 5.5 after the conversion failed with…

    Also about: Hugging Face

    Small models on phones

    A developer publishes two complete Amazon shopping workflows carried out locally by Qwen3.5 0.8B and 2B on Android phones, with a median input of 45 tokens per call; the slower one, on a Galaxy A04e, took 4 h 20 for 578 calls.

  6. Sunday, November 15

    Naxi translation

    Andrew Peters (anthropology, Harvard) fine-tunes Qwen3.5-4B-Base on about 50 000 lines for this language with roughly 300 000 speakers, using a dictionary of 3 619 entries.

    Harel, at the Lexsi Labs hackathon

    A verifier the agent cannot see validates only 4 of the 11 tasks (out of 24) on which it passes the visible tests; after one round of DPO, the pass rate on these tests rises from 45 to 55 % on eight held-out tasks, but validated responses…

  7. Saturday, November 14

    OpenUncensored traffic

    Over one week, 157 paying customers and 150,292 requests: coding agents account for 49% of spending at this uncensored model provider, all agents for 65%, and chat and roleplay for 1.5%, according to the provider.

    CQ and GCMP from ThousandML

    On Qwen3.8-27B, CQ 4-bit quantization reduces the model from 55.6 to 18 GB with no significant difference from BF16 across nine benchmarks, while NF4 loses 12 points on GSM8K; the GCMP adapter matches a rank-32 LoRA with 2.9 million…

    Four posts about Cohere Labs' Aya models

    Fine-tuned on verified corrections, Aya Expanse 32B goes from 14.5% to 51.5% of τ²-bench tasks; Tiny Aya Fire refuses 36 out of 36 harmful requests in Urdu script, but 17 out of 36 in romanized Urdu, according to the authors.

    Also about: Cohere

  8. Friday, November 13

    OpenMed

    Two 73.9-million-parameter extractors, for medications and anatomy, under the Apache-2.0 license; on 200 previously unseen PubMed abstracts, F1 rises from 45.09 to 66.81 and from 26.20 to 61.74.

    TiHon-TTS

    A Vietnamese speech synthesis model with 100,118 parameters (402 KB) runs 40 times faster than real time on one x86 core and can run on a Cortex-M microcontroller, according to its author.

    YOLO26 on NXP i.MX 95

    With its EdgeFirst software, Au-Zone goes from around 40 to 84 frames per second on the 2 TOPS NPU, with a 7-point mAP improvement, according to its measurements.

    s1d1-bench

    An open benchmark compares a multilingual prompt router with three local LLM judges: agreement of 0.807 (QWK), versus 0.774 between judges and 0.278 for RouteLLM's BERT router, according to the author.

  9. Thursday, November 12

    Origami by Wirkung

    Five open models (Qwen3.5-9B and 27B, Gemma 4 31B, MiniMax-H3, Kolibri-1) converted without retraining into files 4.3 to 7.5 times smaller; the blog post provides no measurement of retained quality.

  10. Wednesday, November 11

    shem

    João Galego releases an experimental runtime that runs a Hugging Face model as a container image (a Dockerfile can start with FROM hf.co/...) and, conversely, runs containers on a processor made of model weights; it works “most of the…

    Diseases in Spanish medical texts

    OpenMed releases two open models under Apache-2.0, usable on CPU; according to the team, fine-tuning on DisTEMIST raises F1 from around 36–39 for a generic model to around 73 for the GLiNER2.5 version.

    x402 agent payments

    Left on their own, Qwen3 235B, DeepSeek V4 Pro and Llama 3.3 70B score 59, 55 and 30 points on Fizzl's exam, with none passing; when code decides whether to pay (x402-safe-fetch), their scores rise to 94, 100 and 94.

    GR00T-X-Embodiment-Sim in FiftyOne

    Harpreet Sahota (Voxel51) converts 42 simulated episodes from NVIDIA's dataset, six for each of seven robot bodies, with aligned camera feeds and joint telemetry.

  11. Tuesday, November 10

    GLM-5 on Trainium3

    David Corvoysier runs GLM-5 (754 billion parameters) on a single trn3.48xlarge instance with the published Transformers code: 43 minutes to load, first token in 4.1 seconds, 6.9 tokens per second at a batch size of 1, an initial…

    Also about: Z.ai

    Disease detection at OpenMed

    Maziyar Panahi publishes three open models with 194M to 576M parameters under Apache-2.0 that identify diseases in English medical texts on CPU; fine-tuned on NCBI Disease, they gain about 14 F1 points on BC5CDR.

    Agents that ship without testing

    In a post in Turkish, Serkan Peker shows that 39.4% of 47,293 SWE-agent submissions arrive without testing after the final change, with roughly one-third lower odds of resolution (odds ratio of 0.68); a preregistered hypothesis, replicated…

    Borrowed reasoning

    Timofey Klyubin inserts Qwen3.8-27B's reasoning traces into Qwen3-8B, which rises from 51.1% to 80.9% on 300 LiveCodeBench problems; two 8B models trained to produce these traces do not help.

    EgoDex in FiftyOne

    Harpreet Sahota converts the EgoDex test set released by Apple: 3,243 episodes of 111 manipulation tasks, totaling 7 hours, with poses for 68 joints recorded by an Apple Vision Pro.

  12. Monday, November 9

    Fizzl's x402 exam

    Frits Zwager turns the x402 Agent Safety Bench into an exam open to any agent on Fizzl's website: 19 purchase tasks with simulated sellers, including 14 dishonest ones, paid for in test USDC on Base Sepolia.

    Harness adapters

    A follow-up to the harness-hub project presented on November 6: LoRA training is complete for Liquid AI's LFM2.5-24B and LFM3 and webAI's Twil-8B, in the pi, hermes and opencode harnesses; the team reports up to 43% higher tool efficiency…

    Editing weights for medicine

    Maziyar Panahi (OpenMed) improves small Qwen models' performance on medical exams by modifying their weights without training: Qwen3.5-2B goes from 52.0 to 70.5% on MedQA, with no significant change in mathematics; the gain comes partly…

    Qwen3.8-27B fine-tuned on a Mac Studio

    Roman Melnyk turns Qwen3.8-27B into an assistant for Kafka, Strimzi, Kubernetes and Argo CD, entirely on a Mac Studio M3 Ultra with 512 GB; the 4-bit version fits in around 15 GB, and the author considers this first result mixed.

    Hugging Face and Qai hackathon in Doha

    Aritra Roy Gosthipaty and four Hugging Face coauthors look back on this November 5–7 hackathon, described as Hugging Face's first in the Gulf: more than 500 applications, 150 participants, 40 teams, 15 finalists and 3 winners.

    Also about: Hugging Face

    Humanoid assembling IKEA furniture

    On November 8, Harpreet Sahota converts an excerpt from the 2026 Humanoid IKEA Assembly Challenge for FiftyOne: 6 episodes and 30.7 minutes of a humanoid with 35 motors and seven cameras, with motor torque synchronized with the wrist…

    Meshy's SAM3D-Part and BranchTokens

    SAM3D-Part extracts a complete part from an existing 3D object with one click, with a Chamfer distance around 72% lower than OmniPart across 500 objects (paper accepted at SIGGRAPH Asia 2026, weights released); BranchTokens, developed with…

  13. Sunday, November 8

    HeliosLM

    Chien-Hsin Lin releases an initial series of open models designed for Taiwan: a chat model and a safety classifier derived from Gemma 4 E4B-it, and two speech recognition models derived from Whisper large-v3.

    Shruti

    Udgam Labs, the creator of the small Laghu model, releases 22 speech recognition models under Apache-2.0, one for each language in the Eighth Schedule of the Indian Constitution.

    The Turkish tax

    Serkan Peker estimates that, using vLLM on an H200, producing the same text in Turkish requires 1.45 to 1.62 times the GPU time needed for English across four open models (Turkish-Gemma-9b-T1, Gemma 4 12B, Ministral 3 14B and Qwen3.8-27B),…

    Erasing a style by modifying weights

    Javad Taghia publishes an experiment in which GPT-6 Astra, given tool access to the weights of a small diffusion model (segmind/tiny-sd) but no gradients, must erase Van Gogh's style from it.

    SEW Multimodal AMR in FiftyOne

    On November 7, Harpreet Sahota publishes a FiftyOne conversion of this dataset of 3,151 annotated images across 55 MCAP episodes, captured by SEW-EURODRIVE's autonomous mobile robot using six types of sensors (RGB and thermal cameras,…

  14. Saturday, November 7

    Construction AI Leaderboard

    ONESTRUCTION publishes a model ranking for construction, averaging five benchmarks (plans, BIM models, cost estimation, IDS generation): Claude Opus 5.5 leads with 73.7 points, one tenth ahead of Gemini 4 Argon (73.6), followed by Claude…

    kniv v7.1

    Dragonscale says it improved its six extraction encoders, introduced on November 2, without any retraining: in Claude Code, Claude Opus 5.5 identified a limit of 50 antecedents for coreference, raised to 100 in the ONNX graph; coreference…

    HapTile in FiftyOne

    Harpreet Sahota (Voxel51) converts the HapTile dataset for FiftyOne: 1,699 episodes and 12.36 hours of a teleoperated UR5e arm, with gel-based tactile sensors, two RGB-D cameras, and the level of haptic feedback felt by the operator, under…

  15. Friday, November 6

    Coding agent fleet

    Tony Dzi reviews 103 days of coding agents from the same family running on seven machines: a coordination protocol can satisfy all its safety invariants and still be abandoned.

    Lebanese dialect

    In the second installment of his series, following the one published on 5 November, Bilal Zaraket builds a Lebanese lexicon validated by three native speakers (429 entries) before generating speech data: the share of sentences requiring…

    SuperCLUE ranking

    A VIDRAFT community post shares the November ranking (19 models): GPT-6.1-Sol leads in agentic coding (69,23), and DeepSeek-V4.1-Flash-1017 ranks first among open models in that category (65,38), with Kimi-K3.1-Pro achieving the highest…

    LingYu Robotics

    The company releases 139,4 hours and 13 263 successful episodes of bimanual teleoperation filmed from the robot's viewpoint, across 15 task families, in LeRobot format and under a noncommercial license.

    Egocentric-EMG-Force

    Harpreet Sahota converts eight episodes and 38,2 minutes of household tasks captured by a depth camera for FiftyOne, with an EMG wristband on each wrist recording forearm muscle signals.

  16. Thursday, November 5

    Sand3D

    Sandflow releases 106,094 3D objects, including 10,456 articulated objects, each written as a short Python program in its open sand3d library, with an OpenUSD export viewable in the Hub's 3D viewer.

    SimReady Central

    Lightwheel details its role in NVIDIA's 3D object library on Hugging Face, which now exceeds 3,500 OpenUSD objects: the company validates and tunes their physical behavior during contact and motion.

    DreamTac

    A post introduces this dataset from Peking University, converted by Voxel51 for FiftyOne: 703 trajectories of a Franka arm with gel fingertips, across 12 tasks and 7.03 hours, where contact becomes visible in the images.

    Speech synthesis in Lebanese

    Bilal Zaraket cloned the voice of a tester from Beirut with four models: in blind ratings by five native listeners, the real voice scores 4.98 out of 5, while the best model, VoxCPM2, scores 3.89 and is perceived as Lebanese in only 65% of…

    Guardián colombiano

    Hugo Alonso fine-tunes Tiny Aya Global (3.35 billion parameters) to moderate a public wall of collective memory in Colombian Spanish, on a 6 GB GPU, replacing a chain of three cloud services.

    Anvaya-1B

    A developer pretrains a model with just over one billion parameters from scratch on two 16 GB T4 cards, using PyTorch FSDP and FineWeb, and releases the first checkpoint and the recipe.

  17. Wednesday, November 4

    Synthetic voice detection

    Two researchers at the University of Melbourne release VeriSpoof, a 2.9-billion-parameter audio LLM that detects synthetic voices, attributes their source, and explains its verdict; according to the authors, its VeriSpoof-Plus variant…

    Dharma-AI's Field Extractor

    A 2B model that returns a document's fields as typed JSON; on its in-house benchmark of 546 documents, 0.834 versus 0.816 for Claude Opus 5, compared against models that are not their developers' latest releases, and with no weights found…

    Cribl's OTAP

    A metric that evaluates an agent on the structure of its plan, aligned with reference plans through optimal transport; according to the authors, it agrees more closely with human judges than ROUGE-L across 236 trajectories (0.74 versus…

  18. Tuesday, November 3

    Qwenguard

    In a community post, João Galego turns a RISC-V processor compiled into layers ahead of Qwen3.5-4B into a guardrail that rejects tool calls outside an allowlist: in his tests, no prohibited calls got through and no ordinary requests were…

  19. Monday, November 2

    Dragonscale's kniv v7

    Six open encoders with 55M to 538M parameters (Apache-2.0, ONNX) that extract entities, relations, and facts according to a schema specified at call time, without an LLM, for agent memory; on Re-DocRED, 0.749 within the same sentence and…

    Sliding window and Bifocal layer

    In a community post, Lukas distills Qwen3-4B-Instruct into four student models with equal memory budgets: the sliding window is sufficient for short answers, but on AIME and HMMT his Bifocal layer (a small window and mLSTM state) solves…

    Backdoor blocked by ViKernel

    VIDRAFT implants a backdoor in Qwen3-0.6B through fine-tuning: the word “payroll” in a file path triggers the malicious command in 98.3% of cases on previously unseen paths, without any of 24 checks detecting it; on ViKernel, all 290…

    Beam routing and test cheating

    A preregistered study by Rairo Mukamuri fails to predict test tampering 500 tokens in advance from expert routing in Beam-1.0-501B-A23B: AUROC of 0.580 against a preset threshold of 0.75, for $9.03 in compute.

    LTX 2.5 on an RTX PRO 6000

    GPU rental provider EcoHash renders a 10-second 1080p video with sound in 90 seconds, and in 14.8 seconds at 480p, on a single 96 GB card; the provider's measurements on its own service.

  20. Sunday, November 1

    Pramaan

    In a community post on Hugging Face, Kshitij Thakkar presents a benchmark of 72 geometric pattern tasks inspired by Indian traditions, checked by a verifier.

    AtMem routed-v4 on DolphinBench

    In a community post, Javad Taghia measures his agent memory library at 479 successful tasks out of 600 (79.83%) on this Mem0 benchmark, compared with 390 in the previous run.

    ViKernel

    In a community post, VIDRAFT (FINAL-Bench) presents a v0.1 demo that would embed an agent's controls, including its stop button, in a model's weights.

  21. Saturday, October 31

    Fast Forced Aligner

    A community post on Hugging Face presents MIT-licensed CUDA kernels that identify when each word in a transcript is spoken: 30.44 ms versus 6.34 s for k2 and 28.18 s for NeMo across ten workloads, or about 208 and 926 times faster,…

    Ready-to-use robots

    A community post by projectsim on Hugging Face makes the case for simulation to train and test robots, drawing on third-party results such as Figure's Helix 2.5, which rose from 9% to 56% of tasks completed in 30 previously unseen homes;…

    x402 payments and MetaMask

    A community post by Fizzl on Hugging Face shows MetaMask displaying “high-risk approval” for two $0.001 x402 payments requested by an agent; 289 of the 300 x402 payment addresses sampled on Base fall into the flagged category.

  22. Friday, October 30

    Lumma-Fev-v2

    FrontiersMind releases its 4B and 27B decision models under Apache 2.0: an 87.3% average across eight public benchmarks for the 27B and 82.4% for the 4B, compared to 81.2% for Jev 1.13.0, which retains the edge on BBH and JudgeBench,…

    Qwen3.8-27B on a 16 GB Mac

    Project Pyrite downsizes the model from 51.75 to 10.11 GiB and runs it at 12.2 tokens/s on an M2 Pro with mlx-lm; allocating 3 bits to the 30 most sensitive projections reduces KL divergence by 22% at an equivalent size, according to the…

  23. Thursday, October 29

    Maya 2 Native

    Maya Research releases a speech synthesis model under Apache-2.0 for eleven Indian languages and English, with 20 voices, streaming and cloning from 3 to 10 seconds of audio; according to its measurements, it supports 224 concurrent…

    Granite Switch 4.1.1

    IBM releases a backward-compatible patch: only the transcription model, now multilingual across 10 languages with punctuation, and the decision adapter change, with no improvement figures published.

    Also about: IBM

  24. Wednesday, October 28

    Project Pyrite

    In a community post, a 31.8-billion-parameter MoE, llm-jp-4-32b-a3b-base, shrinks from 59.06 to 9.81 GiB and runs at 34 to 49 tokens per second on a 16 GB Mac M2 Pro, without the repetition loops caused by uniform 2-bit quantization,…

    My Name is Riemann

    OpenPeerAI releases a preview of a 13.1-billion-parameter model obtained by progressively expanding a 163-million-parameter model, on AMD hardware; the authors have run no general benchmarks and flag corpus errors themselves.

    L2Loop

    Yuxuan Zong proposes rerunning certain quantized layers as long as they fit in the GPU's L2 cache: on an RTX 5080, this second pass costs about half as much as a new layer and reduces loss by 0.05 to 0.08 at the same parameter count,…

    EEG datasets

    A community survey of the Hugging Face Hub finds 368 brain-signal datasets, including 146 without a license label, complicating their reuse.

  25. Monday, October 26

    PHP2Shell exploited on honeypots

    Catch-up from October 24: according to The Hacker News, Perplexity researcher Adel Karimi captured what appears to be the first successful exploitation of PHP2Shell (CVE-2026-109618 and CVE-2026-109619) in the wild on his decoy systems…

    Reverse engineering with small models

    In a post with preliminary results, Younes B., Nathan Habib, and Lysandre give two toy binaries to Qwen3.5-9B models guided by DeepSeek-V4.1-Flash-1017 and checked by an oracle with around 600 cases; the lead model intervenes on only about…

    AutoBench's router benchmark

    According to Peter Kruger's measurements, none of the 17 configurations across seven LLM routers beats a fixed call to GPT-5.6 Luna on both quality and cost, and five configurations behave as constants.

    VKORE-325M-NPU-Base

    VIDRAFT pretrains a 325.6M-parameter model from scratch on a commercial inference NPU, using 5 billion FineWeb-Edu tokens: a loss of 0.7993, compared with 0.7994 for the same training on B200; weights are public.

    DecisionFormer

    FrontiersMind releases the code used to train its Lumma-Fev models under Apache 2.0: it turns a Hub model into a decision model that scores options in a single pass, configured through a single YAML file (full fine-tuning, LoRA, or QLoRA).

  26. Sunday, October 25

    Mundo

    Catch-up from October 23: in a Hugging Face post, Yihang Luo, Chen Change Loy and their coauthors introduce a world model that reconstructs a room from a few photos, simulates its objects and hands rendering over to a video model: changing…

    A5

    Catch-up from October 23: Honglin Guo, Tao Gui (SII) and their coauthors introduce a proactive personal assistant that reuses past preferences and decisions, monitors connected sources and suggests next steps on its own.

    Kos

    Mateusz Piesiak releases a Polish voice assistant that is allowed to answer “I'm not sure”: in a Hugging Face Space on ZeroGPU, with no pay-per-use API, Whisper large-v3 listens, the open model PLLuM-12B responds, the code itself handles…

  27. Saturday, October 24

    Beetle

    Researchers from Cambridge, EleutherAI and Oxford publish 330 small models (285 bilingual, in English and one of 21 languages, and 45 monolingual) with checkpoints, data and code, trained on 100M, 2B and 24B tokens according to the post,…

    VocabX

    Catching up on October 21: by rebuilding the vocabulary of its Qwen3-4B-based Portuguese essay grader around the cost of each token, Dharma-AI reduces measured inference cost by 77 % and raises its quality score from 0.703 to 0.810;…

  28. Friday, October 23

    Maverick 1.1

    Manisa Celal Bayar University's XRLab updates its local model with 4 billion parameters that controls Unity scenes in extended reality: it accepts corrections, understands gaze and pointing, acts in other rooms and refuses before acting.

    Visual search on ViDoRe V3

    According to a community post, a pipeline with open weights in which GLM-5.3-Flash drives the EVIE-4.5B retriever achieves an average nDCG@10 of 70.01, topping the ViDoRe V3 pipeline leaderboard on a single GPU.

  29. Thursday, October 22

    Seb-9B-v2

    This local decision model with 9 billion parameters, under Apache 2.0, scores 59,16 on the Decision Index against 57,96 for hosted Jev 1.13 according to its author, who does not specify the index version, taking 22 ms per decision on an…

    Laghu

    UDGAM LABS trains a 152-million-parameter conversational model from scratch for 146 dollars, running at 163 MB in the browser and matching SmolLM2-135M-Instruct in conversation according to its authors.

    BaseQuant V1

    basecompute releases quantizations of Qwen3.8-27B (10,93, 16,41 and 20,82 GB) with 15 % lower KL divergence than Unsloth's at 3 and 4 bits, according to its authors.

    A VLM as a reranker

    Without reranking training, Qwen3.8-27B beats the supervised Qwen3-VL-Reranker-2B reranker on three benchmarks and matches the 8B version on MMEB-V2's visual documents, according to a community post.

Earlier months

October 2026 99
September 2026 139
August 2026 29
July 2026 11
June 2026 5
May 2026 2
April 2026 4
March 2026 2
February 2026 1
January 2026 1