Skip to content
Thursday, 19 November 2026
Tenesys AI News
Subscribe

AI Model Updates

What exactly changed in AI models: context, prices, capabilities. Only values stated by sources — never guessed.

ImportantVoice AIHugging Face · 21 Oct 2026

IBM refreshes Granite Speech lineup: new ASR model, quicker edge transcription, sharper speaker attribution

ParameterBeforeNow
granite-speech-5.0-2b WERNot specified (succeeds granite-speech-4.1-2b)5.06% average English WER, #2 open-weight ASR on leaderboard
Turbo CTC speed (RTFx)~12,600 RTFx at launch19,870 RTFx (roughly 1.6x faster at same accuracy)
Turbo CTC languagesEnglish onlyEnglish, Japanese, Spanish, German, French, Italian, Portuguese
4.1-2B-Plus WDER/cpWERWDER 4.40, cpWER 30.50WDER 4.10, cpWER 28.48 (7% drop each)
4.1-2B-Plus long-form WER13.5310.22 (24% lower)
4.1-2B-Plus keyword biasing WER (20-term)11.796.48 (45% lower)
4.1-2B-Plus timestamp shift (AAS, all speech)210.8698.54
Read more →
ImportantModelsCohere · 21 Oct 2026

Cohere releases Rerank 5 with passage-level retrieval and calibrated scores

ParameterBeforeNow
Context Focus (passage-level retrieval)Not specifiedNew feature: single Rerank call returns scored text spans within documents, not just document-level scores
BRIGHT benchmark score (reasoning-intensive retrieval)Rerank 4 Pro: 69.8 (82.2 minus 12.4 gain)Rerank 5 Pro: 82.2
Relevance score calibrationUncalibrated scores, model-specific thresholdsCalibrated probability scores: 0.7 means ~70% chance document is useful
Throughput (Fast vs Pro)Not specifiedRerank 5 Fast: ~187 docs/sec on A100, 2.9x throughput of Pro, ~3x faster response times
PricingNot specified$0.05 per 1M tokens (Pro), $0.02 per 1M tokens (Fast)
Read more →
ResearchNVIDIA Developer · 21 Oct 2026

NVIDIA and University of Washington Launch RFD4-Proteina for Faster AI-Driven Protein Design

ParameterBeforeNow
Sampling steps~400 (Proteina-Complexa)60
Generation speed vs RFD31x3.8x faster at 200 residues
Peak memory vs RFD3Baseline84% lower at 200 residues
Sequence/structure designSeparate inverse-folding step (e.g. ProteinMPNN) requiredSequence and all-atom structure co-designed in one model
Backend speedup (BioIR vs default)1x1.37x-1.73x across tested Hopper GPUs
Deployment optionsNot specifiedOpen model on GitHub or NIM microservice with HTTP interface
Read more →
ImportantModelsHugging Face · 21 Oct 2026

NVIDIA expands Nemotron with new retrieval, reranking and document-parsing models for AI agents

ParameterBeforeNow
LineupNemotron 3 Embed / Llama Nemotron Embed VL 1BNemotron 3.5 Embed (8B BF16, 1B BF16, 1B NVFP4), Nemotron 3.5 Rerank 8B BF16, Nemotron 3.5 Parse 631M
ViDoRe v3.1 dense retrieval score52.23 nDCG@10 (Llama Nemotron Embed VL 1B)61.17 nDCG@10 (Nemotron 3.5 Embed 8B BF16), rank 1
Reranking score59.17 nDCG@10 (Qwen-3-VL-Reranker-8B)60.49 nDCG@10 (Nemotron 3.5 Rerank 8B BF16)
Agentic token costBaseline embedding-only retrieval~13-14% lower with reranking added, plus higher accuracy
NVFP4 throughput101.3 docs/sec (1B BF16 on GB200)128.8 docs/sec (1B NVFP4), 99.85% of accuracy retained
NIM serving throughput2,099 req/sec (vLLM)6,031 req/sec (NIM, RTX PRO 6000 Blackwell)
Read more →
ImportantModelsAWS Machine Learning · 21 Oct 2026

AWS and i14 ditch the KV cache with a 3B linear-attention model on Trainium2

ParameterBeforeNow
ArchitectureTransformer softmax attention with growing KV cacheGated Linear Attention (GLA) with fixed-size recurrent state
ParametersNot specified3.028B (2.718B Muon + 309.6M AdamW)
Per-sequence memory stateKV cache grows with context lengthFixed 29.5 MB recurrent state regardless of context length
Training throughput73,378 tokens/sec (gather-broadcast Muon)119,934 tokens/sec (all_to_all Muon), ~120,000 tokens/sec overall on trn2.48xlarge
Inference throughputNot specified11,544 output tokens/sec across 1,280 concurrent sequences
Context window2,048 training sequence lengthTested up to 8,192 tokens with only 0.5% extra cost per output token
Read more →
ResearchHugging Face · 20 Oct 2026

NVIDIA Unveils Cosmos3-H-Dreams, a Real-Time Generative Simulator for Surgical Robotics

ParameterBeforeNow
Base modelCosmos-Predict2.5 (dVRK tabletop suturing)Cosmos3-H-Embodied-Nano (16B params, clinical cholecystectomy)
Training dataTabletop suturing video/kinematicsCMR Surgical Versius clinical cholecystectomy recordings
Inference stepsNot specifiedFour-step and two-step distilled student checkpoints
HardwareNot specifiedRuns interactively on a single NVIDIA RTX PRO 6000 GPU
Action dimensionsNot specified28 action dimensions (dual-arm pose, gripper, energy, thumbstick controls)
Training resolution/frequencyNot specified832x480 video, actions aligned at 10 Hz
Read more →
ImportantCyber SecurityAnthropic · 12 Oct 2026

Anthropic cuts Claude distillation by 90-99% with new reasoning-protection defenses

ParameterBeforeNow
Reasoning transcript accessFull chain-of-thought reasoning could often be extracted via active prompting or proxy accountsReasoning summarized, access restricted to originating organization, and new API accounts (post Aug 31, 2026) cannot edi
Distillation detectionSimple summarization-based defenses, bypassed by active extraction techniquesTwo-stage classifier (constitutional classifiers approach) with <0.1% false positive rate, matching bio/cyber safeguard
Extracted reasoning tokensBaseline high volume harvested by distillation networks10-100x lower (some days too low to measure) after full rollout
Read more →
ImportantVoice AIMistral AI · 20 Oct 2026

Mistral launches Voxtral 3 voice stack: speech models plus a real-time agent SDK

ParameterBeforeNow
Languages (STT)Not specified50+ languages including 24 EU languages
STT latency (realtime)Not specifiedSub-500ms in 35+ languages
Mini Transcribe priceNot specified$0.0035/min, 4B parameters
TTS languages/dialectsNot specified35+ languages, 50+ dialects
TTS priceNot specified$13 per 1M characters
Voice cloningNot specifiedPro Voice adaptation from 30 minutes of audio
Tool callingNot specifiedNew Voice Agent SDK combines listening, reasoning, tool calling and speaking in real time
Read more →
ImportantModelsGoogle Developers · 20 Oct 2026

Google launches native PyTorch inference engines for vLLM and SGLang on TPU

ParameterBeforeNow
Serving frameworkJAX-based vLLM/SGLang backends on TPUNative PyTorch backends (TorchTPU) for vLLM and SGLang on TPU
Cost per million tokensNot specified2.01 cents at high-concurrency configurations (lowest reported in SemiAnalysis InferenceX benchmark)
Agentic/long-context supportNot specifiedAutomatic prefix caching, KV cache offloading via TPU Sync, tool calling, structured output, disaggregated prefill/decod
Tokens per second (single user)Not specified700+ tokens/sec for 2.8T-parameter Kimi K3 model on Ironwood TPU via Pallas megakernel
Read more →
ImportantAgentsAWS Machine Learning · 20 Oct 2026

Amazon Bedrock Enforces Schema-Accurate JSON for All Claude Models

ParameterBeforeNow
Output formatFree-form text requiring custom parsing and retriesConstrained decoding guarantees JSON matching a defined schema
Tool callingTool parameters not strictly validated against schemaStrict tool use enforces exact schema match for tool inputs
AvailabilityNot specifiedAll Claude models on Bedrock runtime and Mantle endpoints where the model is offered
Batch/cross-region supportNot specifiedSupported with batch inference and cross-Region inference on the runtime endpoint
Cost impactNot specifiedFewer output tokens since responses only include schema-defined fields, potentially lowering inference cost
Read more →
ModelsAWS Machine Learning · 19 Oct 2026

AWS Bedrock adds TwelveLabs' Pegasus 1.5 for structured video search

ParameterBeforeNow
Max video length1 hour (Pegasus 1.2)2 hours per request
Output formatSingle text answer per clip queryStructured, timestamped JSON segments matching defined criteria
Input methodsPre-indexed clipsDirect S3 URI or base64 string, no pre-indexing
PromptingText-only queriesText plus reference image (multimodal) prompting
APIsNot specifiedInvokeModel, InvokeModelWithResponseStream (US West/Oregon)
Read more →
ImportantVoice AIAWS Machine Learning · 19 Oct 2026

Deepgram cuts batch transcription costs by porting Nova-3 to AWS Trainium2

ParameterBeforeNow
Hardware targetStandard PyTorch on prior compute optionsAWS Trainium2 (trn2.3xlarge) via Neuron/TorchNeuron
Throughput per dollarBaseline compute options Deepgram previously ran11-70% more audio processed per dollar
Serving runtimePython multiprocessing/multithreaded data-parallel runtimeCustom Rust runtime, 7.5% faster than Python MP baseline at max batch size
Replicas per deviceNot specified8 model replicas (DP8) on one trn2.3xlarge instance
Read more →
ImportantVoice AIOpenAI API Changelog · 16 Oct 2026

OpenAI makes custom voice generation available to every GPT-Live API user

ParameterBeforeNow
Voice creation accessLimited availabilityAvailable to all API customers
Voice creation costNot specifiedFree
Speech generation billingStandard GPT-Live session billingStandard GPT-Live session billing (unchanged)
Built-in voicesExisting set+9 new voices: German (Nuri, Juni, Flitz), French (Brise, Sillage), Hindi (Nira), Japanese (Harema, Shida), Korean (Noeu
Read more →
ImportantModelsHugging Face · 16 Oct 2026

Reflection releases Beam, a 501B open-weight MoE model with 1M-token context

ParameterBeforeNow
Total parametersNot specified501B (MoE)
Active parameters per tokenNot specified23B
Context windowNot specified1,000,000 tokens
LicenseNot specifiedApache 2.0
Tool callingNot specifiedSupported via dedicated grammar and response template
Reasoning effort levelsNot specifiedlow, medium (default), high, xhigh, max
Checkpoint formatsNot specifiedBF16, FP8, NVFP4
Read more →
ImportantModelsHugging Face · 15 Oct 2026

IBM combines transcription, decision-making and chat in a single Granite Switch model

ParameterBeforeNow
Capabilities per endpointSeparate models/endpoints for chat, decisions, and speechChat, decision-making, and ASR transcription unified on one endpoint
VRAM usageNot specifiedApproximately same VRAM as base model alone, despite added capabilities
Decision adaptersNo Decision aLoRA in prior releaseGeneral-purpose Decision aLoRA plus specialized Guardian aLoRA for safety decisions
ASR componentNot specifiedGranite-Speech-5.0-Turbo-CTC embedded, top-ranked on OpenASR Leaderboard
Inference speed (decision throughput, 3B demo)Competitor Jev reported at 10 decisions/secondGranite Switch 3B variant reached 60-100 decisions/second
Read more →
ImportantModelsHugging Face · 15 Oct 2026

Meta releases Muse Spark Open, an 816B-parameter open-weight multimodal MoE model

ParameterBeforeNow
ArchitectureMuse Glimmer: dense feed-forward network per layerMuse Spark Open: Mixture-of-Experts, 256 experts per layer, 8 selected per token
Model sizeNot specified~816B total parameters, ~42B active per token
Context windowNot specifiedUp to 1,000,000 tokens
InputsNot specified (Muse Glimmer: no audio encoder)Text, images, audio, and video
OutputNot specifiedText
Checkpoint formatsNot specifiedBF16 and NVFP4 (4-bit, for NVIDIA Blackwell GPUs)
Tool callingNot specifiedSupported, text-only and multimodal
Read more →
ImportantModelsHugging Face · 15 Oct 2026

Apertus 1.5 adds image/audio understanding, longer context and a thinking mode to the open 8B/70B models

ParameterBeforeNow
Context windowNot specified (Apertus 1.0 baseline)262,144 tokens (4x increase over Apertus 1.0)
Input modalitiesText onlyText, images, and audio (audio understanding experimental); generation remains text-only
ReasoningNo explicit reasoning modeOptional 'thinking mode' that reasons before answering, boosting performance on tasks like AIME
Tool callingStandard tool useImproved, trained on execution-verified tool-use trajectories plus RL/DPO tuning
Training data volumeApertus 1.0 pretraining corpus8B model: 19T total tokens (+4T); 70B model: 17T total tokens (+2T)
LicenseApache 2.0Apache 2.0 (unchanged, fully open weights/data)
Read more →
ImportantCyber SecurityGitHub · 13 Oct 2026

GitHub swaps LLM-based secret scanner for a lightweight ModernBERT classifier

ParameterBeforeNow
ArchitectureRemote LLM-based classification pipelineTwo-stage pipeline: regex candidate generation + fine-tuned ModernBERT classifier (ONNX, 8-bit quantized)
PrecisionBaseline (previous LLM pipeline)13% higher (offline evaluation)
RecallBaseline (previous LLM pipeline)19% higher (offline evaluation)
Owner-weighted acceptanceBaseline (previous LLM pipeline)11.5% higher (online A/B)
Median post-push scan execution timeBaseline (previous LLM pipeline)86% lower
Serving costBaseline (previous LLM pipeline)More than 98% lower
Read more →
ImportantModelsHugging Face · 13 Oct 2026

Liquid AI's new open-weight model runs agentic reasoning locally with just 2B active parameters

ParameterBeforeNow
Context window32K tokens131K tokens
Vocabulary size65K tokens128K tokens
Pre-training tokens~17T tokens~33T tokens
ReasoningNot reasoning-onlyReasoning-only with explicit chain of thought
Active parameters2B (predecessor LFM2-24B-A2B)2B (unchanged)
Tool callingNot specifiedSupported, trained inside agent harnesses (OpenClaw, Hermes Agent, Pi, OpenCode)
Inference speed (Apple M5 Max)Not specified184 tok/s
Inference speed (NVIDIA B300)Not specifiedUp to 21K output tok/s
Read more →
ImportantResearchAmazon Science · 12 Oct 2026

Amazon's 'Priming' Turns Pretrained Transformers Into Efficient Hybrid Models for Under 1% of Training Cost

ParameterBeforeNow
Training cost vs scratch100% (full pretraining)Less than 1% (under 0.5% of token budget)
Throughput1x (source Transformer)Up to 2x at 128,000 tokens context
KV cache sizeFull attention cacheHalf the KV cache at inference (32B GKA hybrid)
Context length testedNot specified128,000 tokens
LicenseNot specifiedApache 2.0
Read more →