Radar

Falcon ASR's language support needs language-specific evidence

TII's announcement emphasizes Arabic and Emirati speech, while Portuguese support is listed without a corresponding evaluation result.

HPHelena Prado1 min read

AutoSynthData makes the verifier part of the training example

ServiceNow's pipeline turns capability gaps into executable tasks, checking both successful trajectories and incorrect outcomes.

RMRafael Moura1 min read

Open TTS Leaderboard measures different parts of speech quality

Intelligibility, speaker similarity and playback delay offer distinct evidence; none replaces listening for naturalness and expression.

HPHelena Prado1 min read

MCP · Radar

ProvenanceGuard checks whether the citation belongs to the claim

Multiverse Computing's approach preserves tool-source identities to detect facts attributed to the wrong evidence.

RMRafael Moura1 min read

Beam: sparse computation still needs room for the weights

Reflection's open model separates active computation from total storage, a distinction that matters before deployment.

HPHelena Prado1 min read

Apertus 1.5 makes evaluation conditions part of the release

Multimodal inputs arrive alongside intermediate checkpoints and structured records of how evaluations were run.

HPHelena Prado1 min read

Muse Spark Open: video input starts with frame selection

Meta's multimodal model accepts recordings, but sampling choices determine which visual events reach it.

HPHelena Prado1 min read

FLUX.3 lets prompts specify where objects belong

Region coordinates add placement instructions to generation and source-to-destination control to reference edits.

DSDiego Sato1 min read

LFM2.5 on-device: deciding an action is only the handoff

Liquid AI documents local agent integration; the application still needs to own permissions and tool execution.

RMRafael Moura1 min read

VocabX separates cheaper output from compressed decisions

Dharma-AI's essay-grading example makes vocabulary adaptation a question of decision quality as well as token count.

HPHelena Prado1 min read

TorchNeuron: trace an operation all the way to hardware

Amazon's RMSNorm walkthrough connects profiler totals to device execution, giving optimization work a concrete starting point.

LFLucas Ferraz1 min read

TorchTPU brings a native device path to PyTorch

Transformers and Diffusers can target tpu through familiar APIs, while full-pipeline compatibility remains the engineering check.

LFLucas Ferraz1 min read

AnyLanguageModel gives Swift apps a shared model interface

Local and hosted models share session APIs, with an important qualification when reasoning history moves between providers.

DSDiego Sato1 min read

PyTorch MPS changes kernels—and fixes image output

The PyTorch 2.15 article covers Metal execution paths, FP8 support and a correctness fix worth checking before throughput.

LFLucas Ferraz1 min read

Inkling 2 Small reaches Tinker with its weight release still conditional

Access for inference and fine-tuning precedes the planned weight release. A forecasting example shows why evaluation design deserves separate scrutiny.

HPHelena Prado1 min read

AnyDecisionModel gives Swift applications typed answers instead of prose to parse

The decision API exposes choices and probabilities. Its session behavior and calibration requirements still need explicit handling.

LFLucas Ferraz1 min read

LFM3 makes modality selection part of model selection

A shared backbone supports distinct text, vision, audio, multimodal, and decision variants. Family-level capabilities do not describe every checkpoint.

HPHelena Prado1 min read

Check Trainium model support before provisioning a machine

The support catalog includes software versions and reproduction scripts. Missing entries and passing results both require careful interpretation.

DSDiego Sato1 min read

ThinkingBox rewards agents for the database state they leave behind

Executable outcome checks become training rewards. The reported gains in repeatability come with limits on coverage and transfer.

RMRafael Moura1 min read

Mistral Large 4: check reasoning settings and licensing before integration

Hugging Face documents deployment options for the text-and-image model, including a reasoning setting that does not switch reasoning off.

HPHelena Prado2 min read

ProtoStar puts scientific data reuse to a practical test

The initiative links open research publishing with challenges designed to reveal whether scientific datasets, models, and code are actually reusable.

HPHelena Prado1 min read

ToyBlocks makes robot learning easier to inspect

Controlled worlds help researchers distinguish reusable skills from familiar task sequences.

HPHelena Prado1 min read

Speech recognition needs an accent breakdown

Open ASR Leaderboard adds conversational evidence that an overall score can conceal.

HPHelena Prado1 min read

Cosmos brings simulation controls into generated video

NVIDIA’s checkpoints target interactive scenarios, where visual plausibility still needs physical validation.

HPHelena Prado1 min read

Phonon puts speech on device—with a licensing choice

The TTS retrospective highlights offline synthesis and different terms across checkpoints.

DSDiego Sato1 min read

Dharma targets structured fields beyond OCR

Typed document output is useful only when the extracted relationships match the application’s contract.

LFLucas Ferraz1 min read

Palette separates slide planning from drawing

IBM’s approach combines a visual language with execution checks, making the workflow more specific than text generation.

DSDiego Sato1 min read

CWE-bench tests security repairs in repository context

A vulnerability benchmark is more useful when functional behavior matters—but it is still not a complete security review.

LFLucas Ferraz1 min read

Transformers MIC shrinks checkpoints without discarding the test

Small architectural fixtures target integration errors that can remain invisible during normal execution.

LFLucas Ferraz1 min read

Granite Switch updates adapters while keeping the base

IBM’s patch targets speech and decision components, with compatibility stated for the previous release line.

HPHelena Prado1 min read

What if the model learned its harness?

Harness Hub explores environment-specific adapters, with evidence tied to the tested model and execution setup.

RMRafael Moura1 min read

LeRobot links training data with live instructions

The documented release expands dataset streaming and language interaction around robotic policies.

HPHelena Prado2 min read

Hermes brings its workspace to a browser through the gateway

The v0.23.0 release centralizes session continuity while leaving web access off by default.

RMRafael Moura1 min read

Claude · Radar

Claude Code hooks: timing changes the control

Documented pre-tool and post-tool events serve different purposes, even when both run a check.

BKBia Kanashiro1 min read

Claude · Radar

Subagents separate context, not usage limits

Claude Code documents focused delegation inside a session, with requests still counted against shared usage.

BKBia Kanashiro1 min read

smolagents makes code an action—and execution a choice

Tool composition becomes more flexible, while the executor still determines access to the surrounding environment.

LFLucas Ferraz1 min read