Entity trail

NVIDIA Nemotron

Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.

Briefing refs

Findings

Edges

Sources

Showing the first 40 findings. More graph evidence exists in the corpus.

Corpus findings

2026-09-11 / arxiv-researcherNVIDIA Opens a Natural-Language-Only IMO 2026 Gold Pipeline on Nemotron 3 Ultra, Weights and Proofs IncludedarXiv 2609.10712 describes a system with no formal prover, tools or internet access. It runs three Nemotron 3 Ultra checkpoints (the GA model plus SFT and RL math specialists) in a generate-verify-refine loop and scored 30/42 at IMO 2026, the gold threshold. NVIDIA published Nemotron-3-Labs-Ultra-Math-SFT and -RL on Hugging Face, put the inference recipe in NeMo-Skills, and added Nemotron-IMO-Bench with 200 new olympiad problems. This is one of the first gold-level test-time-compute pipelines anyone can rerun on open weights.
2026-09-05 / github-pulse-researcherllama.cpp 0.4.0 adds lazy tensor reading and a quantizer RAM cap so big models stop peaking host memory on loadReleased 2026-09-04, llama.cpp 0.4.0 adds `llama_lazy_mode` / `--lazy-mode` for on-demand tensor reading (#27794), a `max_buf_size` quantize parameter capping quantizer RAM (#27795), quantizer row-slab streaming (#27830), and a fix preventing RAM peaking during load (#27483). It also adds sparse flash attention for DeepSeek-V4/GLM and Qwen4exp (#27970), Apple RDMA as an RPC transport (#26421), initial Qwen3.8-Flash-Next and NVIDIA Nemotron-3-Puzzle-75B-A9B support, per-slot server context limits and video input options, on ggml 0.23.0.
2026-09-03 / arxiv-researcherFP4 Pretraining With E5M3 Block Scales Beats NVIDIA's Transformer Engine Recipe and Runs 21% FasterStable 4-bit pretraining is hard because the E2M1 payload covers only a narrow magnitude range; NVIDIA's Transformer Engine recipe compensates with current-tensor scaling, a randomized Hadamard transform and BF16 final layers, all of which add work outside the FP4 matmuls. This work pairs E2M1 payloads with unsigned E5M3 block scales whose wider range permits periodic tensor scaling, applies selective stochastic rounding only to backward gradients, drops the Hadamard transform entirely, and uses FP4 in every eligible internal linear. Pretraining a Nemotron-H 8B for nearly 190 billion tokens, the block-16 recipe finished with lower final-window training loss and lower held-out validation NLL under each method's own quantized-inference policy, and an ablation removing both the Hadamard transform and the BF16 final-block exemption raised measured model-body token throughput 21.2%.
2026-09-03 / rss-researcher100 GRPO steps on a free Colab GPU lifted a 350M model's JSON compliance from 18.0% to 31.9%A Hugging Face post published September 3 fine-tuned LiquidAI/LFM2.5-350M with TRL's GRPO on roughly 500 samples from nvidia/Nemotron-RL-instruction_following-structured_outputs, augmented with 40% code-block instructions and 20% top-level-array conversions. Overall IFStruct pass rate went from 22.6% to 29.7% in 100 steps, with JSON format specifically going 18.0% to 31.9%, on a free-tier Colab or Kaggle GPU with 16GB VRAM. The reward is three functions, JSON validity, field count accuracy and schema validation, weighted [1.0, 0.5, 2.0], which is a directly copyable recipe for making a small local model reliable at structured output.
2026-09-03 / hn-researcherMultiverse Computing Launches Quasar 438B and Claims the Top European Score on Artificial AnalysisReleased September 2, Quasar 438B is Multiverse Computing's flagship reasoning model for enterprise agents and coding, scoring 43 on the Artificial Analysis Intelligence Index, 75.0 on AA-LCR long-context reasoning and 69.3 on Terminal-Bench v2.1, with a 15.3-second response time for 500 tokens. That puts it above Mistral Medium 3.5 at 30 and NVIDIA Nemotron 3 Ultra at 38 among European offerings, and well behind Claude Opus 5 at 63. It supports English and Spanish only and is served exclusively through the CompactifAI API, with no weights or license disclosed.
2026-09-02 / news-researcherCrowdStrike and NVIDIA Launched SafeMind, Security Models Trained in an Offense-Versus-Defense Coevolution LoopAt Fal.Con 2026 in Las Vegas on September 1, Jensen Huang joined CrowdStrike CEO George Kurtz to announce CrowdStrike SafeMind, a family of purpose-built security models and agentic harnesses from CrowdStrike's Cyber Superintelligence Lab. It pairs CrowdStrike's own models with defensive models built on NVIDIA Nemotron in a continuous loop where offensive and defensive models repeatedly challenge each other. A separate expanded agentic MDR collaboration using the NVIDIA Agent Toolkit and NeMo Data Designer claims 5x faster investigations and 3x higher triage accuracy, and NVIDIA's OpenShell open-source runtime is being integrated into the Falcon platform.
2026-09-01 / hn-researcherThe Kuleshov Group Published a Full Build Guide for Diffusion Language Models, With Mercury 2 Claiming 1,000+ Tokens/SecondAdapted from ICLR 2026 and MLSS 2026 workshops, the guide (178 points on HN) walks masked diffusion models, block diffusion for variable-length generation, encoder-decoder architectures, remasking-based error correction, sampling distillation, guidance and RL post-training. It catalogs what is actually available to build on: LLaDA at 8B with open weights and LLaMA compatibility, Google's Gemma Diffusion, NVIDIA's Nemotron Diffusion family up to 35B, ESM3 at 100B for proteins, and Nucleotide Transformer v3 on roughly a trillion DNA tokens. The commercial datapoint is Mercury 2 at 1,000+ tokens/second on standard GPUs, which the authors put at 5-10x the throughput of comparable-quality models like Claude Haiku.
2026-08-28 / thought-leaders-researcherKwindla Hultman Kramer open-weights PhoneLLM, a 30B/3.5B-active voice-agent model claiming GPT 5.6 Terra parity at 1/18 the costThe Pipecat/Daily team released phonellm-alpha-1 under BSD 2-Clause, a MoE built on NVIDIA Nemotron 3 Nano 30B-A3B and tuned for low-latency tool calling and instruction following rather than general chat. The model card claims parity with GPT 5.6 Terra at 94% lower cost and 1,300ms faster P95 time-to-first-token, sub-100ms single-request TTFT on a B200, and 88 concurrent agents per B200 at roughly $0.00025 per agent-minute. The pitch is model-size arbitrage: for phone agents, latency and tool-call reliability beat raw capability.
2026-08-13 / rss-researcherMeta and NVIDIA both dropped open-weight models this week in a coordinated answer to Chinese labsCNBC reported August 12 that Meta unveiled Muse Glimmer on Monday and NVIDIA debuted Nemotron 3.5 Lightning a day later, framing both as a U.S. push against Chinese labs' lead in open weights. Both companies were among 20+ U.S. firms that signed a letter last month urging policymakers against 'premature restrictions' on open-weight models. Box CEO Aaron Levie, a signatory, called Zuckerberg's plan to open the weights for Muse Spark 1.2 a 'very big deal' and said 'there's a very firm flag in the ground that America will have near-frontier open-source models.'
2026-08-12 / agents-researcherNVIDIA ships Nemotron 3.5 Lightning, a 30B MoE built for agent loops, plus NeMo Switchyard routing that cuts task cost to a third of Opus 4.8NVIDIA released Nemotron 3.5 Lightning on August 11 — a 30-billion-parameter mixture-of-experts open model claiming up to 4x faster output and 30% faster agentic task completion versus comparable models, aimed at the high-volume, narrow roles inside multi-agent systems (code review, tool use, security monitoring, billing triage). Alongside it, NeMo Switchyard is an open-source routing library that dispatches each request to the cheapest model that can still handle it; NVIDIA's internal testing claims frontier-level accuracy at roughly one-third the task cost of running Opus 4.8 alone. NVIDIA also published Nemotron-RL-Agentic-Terminal-Pivot, an agentic RL dataset for coding-agent post-training. For builders, Switchyard is the more interesting artifact: it makes model-tier routing a library concern rather than something every agent harness reimplements.
2026-08-04 / reddit-researcherNVIDIA Ships NemotronLabs VoiceChat 11B — First Open Full-Duplex Speech Model With Tool Calling, ~450ms Turn-TakingNVIDIA uploaded NVIDIA-NemotronLabs-VoiceChat-11B to Hugging Face on August 3: a Fast Conformer speech encoder in front of Nemotron Nano v2 9B (hybrid Mamba/Transformer) with an NVIDIA TTS decoder behind it, collapsing the usual ASR→LLM→TTS cascade into one model. The card reports ~450ms latency on smooth turn-taking, 480ms on user interruption, and #2 among open full-duplex models on VoiceBench, released under the OpenMDW 1.1 license. The genuinely new part is a separate output channel that emits tool-call scripts while audio keeps flowing, with configurable 'on-hold' phrases the agent speaks during tool execution — the first open full-duplex model to do this.
2026-07-28 / sources-researcherNVIDIA Turns PhysicsNeMo and CUDA-X Into Agent-Callable Skills, and Claims Nemotron 3 Ultra Leads Open Models on Agentic RTL CodingNVIDIA expanded its Agent Toolkit with re-architected PhysicsNeMo libraries and updated CUDA-X libraries exposed as agent-ready tools and skills, so engineering agents can invoke AI physics models, accelerated solvers and quantum chemistry directly rather than through bespoke wrappers. Alongside it, NVIDIA Research's ACE-RTL agent paired with Nemotron 3 Ultra is claimed to lead open models on agentic register-transfer-level coding across the comprehensive Verilog design problems benchmark. ChipAgents is already fine-tuning Nemotron models on the toolkit for end-to-end semiconductor debug, formal verification and coverage — a concrete example of the 'domain library becomes a tool surface' pattern spreading beyond software.

Graph relationships

BUILT BY
PhoneLLM
PhoneLLM MoE built on NVIDIA Nemotron 3 Nano 30B base
Source finding
BUILT BY
PhoneLLM -> NVIDIA Nemotron
PhoneLLM MoE built on NVIDIA Nemotron 3 Nano 30B base
Source finding

Source trail

Released September 2 serverless fine-tuning for NVIDIA Nemotron 3 on SageMaker semantic layer for agentic AI using Stardog on Bedrock AgentCore four deployment patterns for Unsloth-quantized models M3 scores 59%Unveiled at Computex With Flash already beating 3.1 Pro Identifiers spotted Command A+W4A4 quantization is lossless 76.2% on Terminal-Bench 2.1 The $100 tier

Graph sources

entity graphfindings textkg edgeskg entitiesnewsletter issues

Fetching from the wire…