NVIDIA releases multimodal Nemotron 3.5 Super VL for AWS EKS
| Parameter | Before | Now |
|---|---|---|
| Parameters | Not specified | 120B total, 12B active per token |
| Context window | Not specified | 1M tokens |
| Architecture | Not specified | Hybrid Mamba-2, MoE, and Attention |
What exactly changed in AI models: context, prices, capabilities. Only values stated by sources — never guessed.
| Parameter | Before | Now |
|---|---|---|
| Parameters | Not specified | 120B total, 12B active per token |
| Context window | Not specified | 1M tokens |
| Architecture | Not specified | Hybrid Mamba-2, MoE, and Attention |
| Parameter | Before | Now |
|---|---|---|
| Parameters | Not specified | 1 trillion |
| Reasoning modes | Not specified | Low, Medium, High |
| Terminal-Bench 4.0 | 28% | 41% |
| Agent Security (B3) | 93.3% | 97.1% |
| Parameter | Before | Now |
|---|---|---|
| Granite Speech ASR | Not specified | Multilingual support for 10 languages with native punctuation and capitalization |
| Decision adapter | Not specified | Improved performance on the Jev Decision Index test set |
| Parameter | Before | Now |
|---|---|---|
| Input price | Not specified | $2 per million tokens |
| Output price | Not specified | $10 per million tokens |
| Cached input price | Not specified | 95% off input price |
| Parameter | Before | Now |
|---|---|---|
| Languages (Transcription) | Not specified | 40 |
| Translation support | Not specified | 31 languages to/from English |
| Speaker attribution | Not specified | 13 languages (up to 8 speakers) |
| Parameter | Before | Now |
|---|---|---|
| Time to first audio (median) | Not specified | 47 ms |
| Price | Not specified | $15 per 1M characters |
| Parameter | Before | Now |
|---|---|---|
| Memory enforcement | Not supported | Transparent per-workload limits with burstable maximums |
| Compute modes | Time-slicing only | Time-slicing and sm-sharing |
| Parameter | Before | Now |
|---|---|---|
| Parameters | LFM2.5-2.6B | LFM3-1.7B |
| Context window | Not specified | 256K tokens |
| Pre-training | Not specified | 22T tokens |
| Parameter | Before | Now |
|---|---|---|
| Repository version | Not specified | v0.2.0 |
| Parameter | Before | Now |
|---|---|---|
| Multimodality | Not specified | Native audio, interactive dialogue, and visual world simulation |
| Reasoning agent | Not specified | Gemini 3.8 Flash used for multi-step reasoning and scene steering |
| Parameter | Before | Now |
|---|---|---|
| MoE Backend | FlashInfer MXFP4 | Humming |
| LM Head Precision | FP32 | Online FP8 |
| Speculative Decoding | None | MTP3 (NEXTN layer) |
| Parameter | Before | Now |
|---|---|---|
| Inference format | Direct JSON | Swappable (Express, Atom, Elemental, or custom) |
| Token efficiency (Express vs Direct JSON) | Not specified | Up to 3x fewer tokens |
| Latency reduction (Express) | Not specified | Up to 3x lower |
| Parameter | Before | Now |
|---|---|---|
| Task verification accuracy | Not specified | 91.3% |
| Mean absolute frame distance | Not specified | 0.96 seconds |
| Parameter | Before | Now |
|---|---|---|
| Capability | Not specified | Video generation/editing via text, images, and frames |
| Output Resolution | Not specified | 360p, 720p, 1080p, and 4k (on paid tiers) |
| Max Clip Duration | Not specified | 10 seconds (up to 40 seconds via extensions) |
| Parameter | Before | Now |
|---|---|---|
| Model Name | Not specified | gpt-6.1-terra |
| Parameter | Before | Now |
|---|---|---|
| Input pricing | Not specified | Reduced via cache reuse (up to 62% in tests) |
| Supported Models | Not specified | GPT-6 Astra, GPT-6.1 Sol, GPT-5.6 Luna |
| Parameter | Before | Now |
|---|---|---|
| Voice catalog | 43 generative voices | 57 generative voices across 11 languages and 25 locales |
| New voices | Not specified | 14 new voices across 9 locales |
| New languages | Not available in Generative engine | Hindi and Finnish generally available |
| API | SynthesizeSpeech API, same voice names | Unchanged; drop-in upgrade, no code changes required |
| Pricing | Standard pay-as-you-go pricing | Same price, no cost increase, no new SKU |
| Region availability | 10 AWS Regions | 11 AWS Regions including GovCloud (US-West) |
| Accuracy | Baseline generative voice accuracy | Improved accurate rendering of numbers, currency, addresses, dates |
| Streaming | Not specified | Bi-Di Streaming API (StartSpeechSynthesisStream) supported for real-time voice agents |
| Parameter | Before | Now |
|---|---|---|
| Parameters | Not specified | 320B (18B active, MoE) |
| Context window | Not specified | 1M tokens |
| Input price | Not specified | $0.15/M tokens |
| Cached input price | Not specified | $0.03/M tokens |
| Output price | Not specified | $0.50/M tokens |
| Tool calling | Not specified | Supported, designed for complex tool-use tasks |
| Parameter | Before | Now |
|---|---|---|
| Output | Not specified | One 128-dimensional vector per 10m pixel, derived from a year of Sentinel-1, Sentinel-2 and Landsat time series; a 64-di |
| Benchmark performance | OlmoEarth v1.2 | Tops AlphaEarth Foundations and Tessera v2 on 20 of 27 evaluations (128-dim) and 18 of 27 (64-dim) |
| Training compute | v1 baseline | v1.3 used 53% of v1.2's training time and 18% of v1's, cutting carbon output to 50% and 22% of those versions respective |
| Reproduction cost | Not specified | Roughly $2,720 in rented H100 time for the final training run |
| Licensing | OlmoEarth Artifact License (v1, v1.2) | Same open license applied to v1.3, weights, data and code |
| Parameter | Before | Now |
|---|---|---|
| Inference speed | Baseline Cosmos 3 NIM microservices | Up to 2x faster in validated configurations, no quality loss |
| Model size (MicroAGI case) | 32B-parameter reasoner | Distilled to 8B-parameter Nano model, no accuracy loss |
| Video stream monitoring (Metropolis VSS 3.3) | Baseline GPU throughput | 40% more streams monitored on same GPU |
| Parameter | Before | Now |
|---|---|---|
| granite-speech-5.0-2b WER | Not 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 launch | 19,870 RTFx (roughly 1.6x faster at same accuracy) |
| Turbo CTC languages | English only | English, Japanese, Spanish, German, French, Italian, Portuguese |
| 4.1-2B-Plus WDER/cpWER | WDER 4.40, cpWER 30.50 | WDER 4.10, cpWER 28.48 (7% drop each) |
| 4.1-2B-Plus long-form WER | 13.53 | 10.22 (24% lower) |
| 4.1-2B-Plus keyword biasing WER (20-term) | 11.79 | 6.48 (45% lower) |
| 4.1-2B-Plus timestamp shift (AAS, all speech) | 210.86 | 98.54 |
| Parameter | Before | Now |
|---|---|---|
| Context Focus (passage-level retrieval) | Not specified | New 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 calibration | Uncalibrated scores, model-specific thresholds | Calibrated probability scores: 0.7 means ~70% chance document is useful |
| Throughput (Fast vs Pro) | Not specified | Rerank 5 Fast: ~187 docs/sec on A100, 2.9x throughput of Pro, ~3x faster response times |
| Pricing | Not specified | $0.05 per 1M tokens (Pro), $0.02 per 1M tokens (Fast) |
| Parameter | Before | Now |
|---|---|---|
| Sampling steps | ~400 (Proteina-Complexa) | 60 |
| Generation speed vs RFD3 | 1x | 3.8x faster at 200 residues |
| Peak memory vs RFD3 | Baseline | 84% lower at 200 residues |
| Sequence/structure design | Separate inverse-folding step (e.g. ProteinMPNN) required | Sequence and all-atom structure co-designed in one model |
| Backend speedup (BioIR vs default) | 1x | 1.37x-1.73x across tested Hopper GPUs |
| Deployment options | Not specified | Open model on GitHub or NIM microservice with HTTP interface |