{"data":{"id":"thr_open_weights_race_muse_spark_open_nemotron_beam_","actors":[{"first":"2026-10-15","last":"2026-10-28","mentions":6,"name":"Nvidia"},{"first":"2026-10-17","last":"2026-10-27","mentions":4,"name":"DeepSeek"},{"first":"2026-10-16","last":"2026-10-25","mentions":2,"name":"Epoch AI"},{"first":"2026-10-15","last":"2026-10-15","mentions":1,"name":"Alexandr Wang"},{"first":"2026-10-27","last":"2026-10-27","mentions":1,"name":"AllModels"},{"first":"2026-10-16","last":"2026-10-16","mentions":1,"name":"Anthropic"},{"first":"2026-10-28","last":"2026-10-28","mentions":1,"name":"Artificial Analysis"},{"first":"2026-10-21","last":"2026-10-21","mentions":1,"name":"DeepSeek Harness lead Cui Tianyi"},{"first":"2026-10-21","last":"2026-10-21","mentions":1,"name":"DeepSeek senior researcher Chen Deli"},{"first":"2026-10-22","last":"2026-10-22","mentions":1,"name":"European Commission"},{"first":"2026-10-22","last":"2026-10-22","mentions":1,"name":"France"},{"first":"2026-10-26","last":"2026-10-26","mentions":1,"name":"FrigadeHQ"},{"first":"2026-10-22","last":"2026-10-22","mentions":1,"name":"Germany"},{"first":"2026-10-22","last":"2026-10-22","mentions":1,"name":"Greece"},{"first":"2026-10-21","last":"2026-10-21","mentions":1,"name":"Jensen Huang"}],"briefs":[],"built_at":"2026-10-29T09:16:21.420746+00:00","coverage_series":[{"count":0,"day":"2026-10-24","editions":0},{"count":1,"day":"2026-10-25","editions":0},{"count":1,"day":"2026-10-26","editions":0},{"count":3,"day":"2026-10-27","editions":0},{"count":3,"day":"2026-10-28","editions":0},{"count":4,"day":"2026-10-29","editions":0},{"count":0,"day":"2026-10-30","editions":0}],"current":{"bottom_line":"Nvidia's open Nemotron 3.5 Transcribe and Granary v2 on Oct 28 pushed the open-weights race into voice AI with promptable multilingual transcription, translation and speaker recognition, while Liquid AI's LFM3 on Oct 27 added another specialized multimodal open family. The Artificial Analysis Intelligence Index v5 released Oct 28 reframes the surge by measuring a wider gap between open and closed models, consistent with independent leaderboards showing open bests at roughly half of closed records on several benchmarks. Abliterated DeepSeek V4.1 Flash variants reaching five inference providers by Oct 27 shows how quickly open weights spread to modified distributions, with high confidence on…","coverage":"steady coverage","date":"2026-10-29","escalation":{"direction":"steady","pace":"gradual"},"headline":"Nvidia's open voice model and new index highlight widening open-closed gap despite fresh releases","source":"daily","url":"/ai/2026-10-29"},"days_covered":6,"domain":"ai","escalation_history":[{"date":"2026-10-19","direction":"steady","pace":"flat"},{"date":"2026-10-21","direction":"steady","pace":"flat"},{"date":"2026-10-24","direction":"steady","pace":"gradual"},{"date":"2026-10-25","direction":"steady","pace":"flat"},{"date":"2026-10-27","direction":"rising","pace":"gradual"},{"date":"2026-10-29","direction":"steady","pace":"gradual"}],"figures":[{"label":"Flora cost per hour of robot video labeling (projection at $10/GPU/hour","series":[{"as_of":"2026-10-25","source":"https://the-decoder.com/nvidia-open-sources-flora-to-make-robot-video-labeling-125x-cheaper-than-gpt-6-astra/","value":0.72}],"unit":"$"},{"label":"Flora labeling throughput for GR00T training","series":[{"as_of":"2026-10-25","source":"https://research.nvidia.com/labs/gear/flora/","value":200000}],"unit":"hours per day"},{"label":"OpenMobile-2 27B SPA-Bench score on real Android devices","series":[{"as_of":"2026-10-25","source":"https://the-decoder.com/openmobile-2-proves-open-source-smartphone-agents-can-rival-proprietary-models/","value":59.6}],"unit":"score"},{"label":"Nvidia Nemotron 3.5 Super VL parameters","series":[{"as_of":"2026-10-21","source":"https://blogs.nvidia.com/blog/nemotron-3-5-super-vl/","value":120000000000}],"unit":"parameters"},{"label":"Tenstorrent Series E post-money valuation","series":[{"as_of":"2026-10-19","source":"https://techcrunch.com/2026/10/19/tenstorrent-hits-6-2b-valuation-to-scale-open-weight-ai-chips/","value":6.2}],"unit":"$B"},{"label":"Tenstorrent post-money valuation after Series E","series":[{"as_of":"2026-10-19","source":"https://techcrunch.com/2026/10/19/tenstorrent-hits-6-2b-valuation-to-scale-open-weight-ai-chips/","value":6200000000}],"unit":"USD"},{"label":"DeepSeek-V4.1-Flash DeepSWE v1.1 score","series":[{"as_of":"2026-10-17","source":"https://the-decoder.com/deepseek-v4-1-flash-1017-beats-larger-rivals-in-agentic-tasks-while-once-again-massively-cutting-per-task-costs/","value":77}],"unit":"%"},{"label":"DeepSeek-V4.1-Flash-1017 DeepSWE v1.1 score","series":[{"as_of":"2026-10-17","source":"https://the-decoder.com/deepseek-v4-1-flash-1017-beats-larger-rivals-in-agentic-tasks-while-once-again-massively-cutting-per-task-costs/","value":77}],"unit":"%"},{"label":"Beam total parameters","series":[{"as_of":"2026-10-16","source":"https://reflection.ai/blog/introducing-beam","value":501}],"unit":"B"},{"label":"Beam total parameters","series":[{"as_of":"2026-10-16","source":"https://huggingface.co/blog/reflection-beam","value":501000000000}],"unit":"parameters"},{"label":"Beam total parameters (23B active per token","series":[{"as_of":"2026-10-16","source":"https://huggingface.co/blog/reflection-beam","value":501000000000}],"unit":"parameters"},{"label":"Reflection Beam total parameters","series":[{"as_of":"2026-10-16","source":"https://huggingface.co/blog/reflection-beam","value":501000000000}],"unit":"parameters"},{"label":"Z.ai GLM-5.3 Cyber ECI (top open","series":[{"as_of":"2026-10-16","source":"https://epoch.ai/data-insights/open-closed-cyber-eci-gap","value":160.1}],"unit":"ECI"},{"label":"Muse Spark Open active per token","series":[{"as_of":"2026-10-15","source":"https://huggingface.co/blog/meta-muse-spark-open","value":42000000000}],"unit":"parameters"},{"label":"Muse Spark Open total parameters","series":[{"as_of":"2026-10-15","source":"https://research.meta.ai/blog/muse-spark-open","value":816000000000}],"unit":"parameters"},{"label":"Muse Spark Open total parameters","series":[{"as_of":"2026-10-15","source":"https://research.meta.ai/blog/muse-spark-open","value":816}],"unit":"B"},{"label":"Muse Spark Open total parameters (~42B active per token","series":[{"as_of":"2026-10-15","source":"https://huggingface.co/blog/meta-muse-spark-open","value":816000000000}],"unit":"parameters"},{"label":"Epoch Cyber ECI top open-weight score (Z.ai GLM-5.3","series":[{"as_of":"2026-10-08","source":"https://epoch.ai/data-insights/open-closed-cyber-eci-gap","value":160}],"unit":"score"}],"first_seen":"2026-10-19","forecasts":[],"indicators":[],"inherited":[],"last_novel":"2026-10-29","last_seen":"2026-10-30","lineage":{"absorbed":[],"branches":[],"merged_into":null,"parent":null},"links":[{"date":"2026-10-25","link":"Gemini 4 Argon frontier push and challengers <-> Open-weights race: Muse Spark Open, Nemotron, Beam and on-device models: Kimi K3.1 Pro scoring 55 on Oct 23 to within 3 points of Claude Opus 5.5 at 58 narrows the open-proprietary gap, but with no new independent benchmarks in three days and LMArena/SWE-bench leaderboards not currently updating the measured open-weights gap remains unchanged and the Oct 25 OpenMobile-2 claim remains unverified.","name":"Gemini 4 Argon frontier push and challengers","theater_id":"thr_gemini_4_argon_frontier_push_and_challengers"},{"date":"2026-10-24","link":"Open-weights race: Muse Spark Open, Nemotron, Beam and on-device models ↔ Neocloud financing stress and chip challengers to Nvidia: shared actor Nvidia — showcasing European adoption of open models at GTC Berlin Oct 22 while positioned as the incumbent facing chip challengers amid debate over hyperscaler capex and leverage","name":"Neocloud financing stress and chip challengers to Nvidia","theater_id":"thr_neocloud_financing_stress_and_chip_challengers_t"}],"name":"Open-weights race: Muse Spark Open, Nemotron, Beam and on-device models","on_record":[{"about":["Russia","Poland","Germany"],"about_iso2":["RU","PL","DE"],"affiliation":"Poland","date":"2026-10-19","id":"st_4f98cc24a319","iso2":"PL","paraphrase":"Russia was behind the late-September sabotage of a Starlink satellite communications station in central Poland.","quote":"hybrid warfare","reported_by":"Deutsche Welle","role":"Prime Minister","signal":"accusation","significance":"Polish prime minister directly attributes Starlink sabotage in Poland to Russia and labels it hybrid warfare, escalating attribution narrative.","speaker":"Donald Tusk","stance":-2,"url":"https://www.dw.com/en/bulgaria-seeks-suspects-in-germany-after-starlink-sabotage/a-79706745?maca=en-rss-en-all-1573-rdf","venue":"other"},{"about":["Russia","Germany"],"about_iso2":["RU","DE"],"affiliation":"Poland","date":"2026-10-19","id":"st_c3b4a94ef2fe","iso2":"PL","paraphrase":"Russia cooperated with organized crime elements in Germany on the alleged Starlink sabotage operation in Poland.","quote":"","reported_by":"Deutsche Welle","role":"Domestic intelligence agency","signal":"accusation","significance":"Polish intelligence alleges Russian use of organized crime networks in Germany for sabotage, implicating cross-border hybrid operation.","speaker":"Polish domestic intelligence agency","stance":-2,"url":"https://www.dw.com/en/bulgaria-seeks-suspects-in-germany-after-starlink-sabotage/a-79706745?maca=en-rss-en-all-1573-rdf","venue":"other"},{"about":["France","Netherlands","Europe"],"about_iso2":["FR","NL"],"affiliation":"France and the Netherlands","date":"2026-10-14","id":"st_152fdf0fb102","iso2":"","paraphrase":"At the meeting they discussed strengthened bilateral cooperation and strategic convergence, agreed a way forward, and reaffirmed their ambition for a stronger and safer Europe.","quote":"reaffirmed their ambition for a stronger and safer Europe","reported_by":"","role":"Representatives of the Elysée and the Office of the Prime Minister of the Netherlands","signal":"commitment","significance":"Signals Franco-Dutch strategic convergence on European security and intent to deepen cooperation within the French deterrence framework.","speaker":"France and the Netherlands","stance":1,"url":"https://www.elysee.fr/emmanuel-macron/2026/10/15/france-and-the-netherlands-held-their-first-nuclear-steering-group-meeting-in-paris","venue":"other"},{"about":["European Union","France","Germany"],"about_iso2":["EU","FR","DE"],"affiliation":"European Commission","date":"2026-10-13","id":"st_1df4f01396cd","iso2":"","paraphrase":"The EU has launched 19 AI Factories and 13 national Antennas and is mobilising more than 30 billion euros for up to seven AI Gigafactories, with the Cloud and AI Development Act to triple computing capacity over 5-7 years, and is working with France, Germany and other Member States to build frontier model capabilities.","quote":"","reported_by":"","role":"Executive Vice-President of the European Commission","signal":"policy_announcement","significance":"Details scale of EU sovereign compute infrastructure push to reduce dependence on non-EU frontier models.","speaker":"Henna Virkkunen","stance":0,"url":"https://ec.europa.eu/commission/presscorner/detail/en/speech_26_2150","venue":"other"}],"places":[{"count":3,"country":"Germany","last_date":"2026-10-22","lat":52.52,"lon":13.405,"name":"Berlin"}],"primer":{"built_at":"2026-10-27T09:39:23.283184+00:00","text":"Open-weights models publish their trained weights so anyone can download, run and adapt them, unlike closed models kept behind an API. Families such as Muse Spark Open, Nemotron, Beam and DeepSeek compete to narrow the performance gap with closed systems, while chipmakers like Nvidia provide the hardware and software they run on. Because weights are open, models can be run on-device — on phones, laptops or even game consoles using tools like llama.cpp — and deployed on sovereign infrastructure controlled locally rather than relying on a foreign cloud. That portability is why an open release can quickly spread across inference providers and to modified, or abliterated, variants with safety filters removed."},"pulses":[{"band":"severe","history":[{"at":"2026-10-25T09:15:03.533845+00:00","position":68}],"id":"pls_ai_geopolitics_ai_safety_governance_fracture","name":"AI Safety Governance Fracture","position":68,"situation":"AI geopolitics"},{"band":"severe","history":[{"at":"2026-10-01T03:00:57.470982+00:00","position":65}],"id":"pls_us_china_technology_export_control_decoupling","name":"Technology and Export-Control Decoupling","position":65,"situation":"US–China competition"},{"band":"elevated","history":[{"at":"2026-10-01T03:00:57.470982+00:00","position":35}],"id":"pls_ai_geopolitics_compute_control_erosion","name":"Compute Control Erosion","position":35,"situation":"AI geopolitics"}],"relations":[],"reports":[{"date":"2026-10-29","url":"/ai/2026-10-29"},{"date":"2026-10-27","url":"/ai/2026-10-27"},{"date":"2026-10-25","url":"/ai/2026-10-25"},{"date":"2026-10-24","url":"/ai/2026-10-24"},{"date":"2026-10-21","url":"/ai/2026-10-21"},{"date":"2026-10-19","url":"/ai/2026-10-19"}],"schema":"ohmega.dossier/1","state":"active","statements":[{"quote":false,"role":"Chief AI Officer, Meta","said":"Muse Spark Open is the most capable US open-weight model to date","source":"https://www.straitstimes.com/world/united-states/meta-unveils-new-open-ai-model-to-rival-chinese-competition","when":"2026-10-15","who":"Alexandr Wang"}],"theater_id":"thr_open_weights_race_muse_spark_open_nemotron_beam_","timeline":[{"date":"2026-10-28","detail":"Released Oct 28 as open models combining transcription, translation and speaker recognition that can be prompted and adapted, with Granary v2 dataset. Reported by Hugging Face blog, TechCrunch and The Decoder.","from":"/ai/2026-10-29","headline":"Nvidia released open Nemotron 3.5 Transcribe and Granary v2 on Oct 28","sources":["https://huggingface.co/blog/nvidia/nemotron-transcribe-granary","https://techcrunch.com/2026/10/28/voice-ai-in-vogue-nvidia-releases-new-open-transcription-model/","https://the-decoder.com/nvidias-open-nemotron-3-5-transcribe-combines-transcription-translation-and-speaker-recognition/"],"verification":"reported","where":""},{"date":"2026-10-28","detail":"Intelligence Index v5 published Oct 28 reported as widening the measured gap between open and closed models, covered by The Decoder. No independent replication yet in window.","from":"/ai/2026-10-29","headline":"Artificial Analysis Intelligence Index v5 on Oct 28 widens measured open-closed gap","sources":["https://artificialanalysis.ai/articles/artificial-analysis-intelligence-index-v5","https://the-decoder.com/new-artificial-analysis-index-widens-the-gap-between-open-and-closed-models/"],"verification":"reported","where":""},{"date":"2026-10-27","detail":"Announced Oct 27 as family for building specialized multimodal models covering every modality for specialized intelligence, per Hugging Face and Liquid AI blogs.","from":"/ai/2026-10-29","headline":"Liquid AI released LFM3 on Oct 27 for specialized multimodal models","sources":["https://huggingface.co/blog/LiquidAI/lfm3","https://www.liquid.ai/blog/lfm3"],"verification":"reported","where":""},{"date":"2026-10-27","detail":"As of Oct 26-27, abliterated DeepSeek V4.1 Flash variants (552B total, 16B active, 1M context, unchanged tokenizer) are served by multiple providers via routing layers such as AllModels (single API with automatic fallbacks, modality-aware routing) and OrcaRouter. Abliteration removes refusal direction by orthogonalizing weight matrices with no retraining. There were 35 Hugging Face repos carrying the name as of Oct 26, with varying effectiveness (e.g., refusals 97→24 and 97→4 on chat prompts). No publisher has released capability benchmarks for abliterated builds; cited capability figures were measured on base checkpoint.","from":"/ai/2026-10-27","headline":"Abliterated DeepSeek V4.1 Flash variants spread to 5+ inference providers via routing layers by Oct 27","sources":["https://allmodels.io/llm/abliterated","https://www.orcarouter.ai/blog/deepseek-v4-1-flash-abliterated-explained"],"verification":"researched","where":"global"},{"date":"2026-10-26","detail":"Released Hob on Oct 26 as free MIT-licensed desktop app for Mac, Windows and Linux that runs open-weight models like Gemma and gpt-oss locally, privately and offline, positioned as alternative to Claude/ChatGPT.","from":"/ai/2026-10-27","headline":"FrigadeHQ released Hob on Oct 26 as open-source on-device desktop alternative for local models","sources":["https://github.com/FrigadeHQ/hob"],"verification":"researched","where":"global"},{"date":"2026-10-25","detail":"Single outlet report on Oct 25 claims OpenMobile-2 proves open-source smartphone agents can rival proprietary models. No paper, benchmark table, model weights link or independent test was included in the evidence provided, so performance, tasks and device constraints remain unverified.","from":"/ai/2026-10-25","headline":"OpenMobile-2 reported on Oct 25 to rival proprietary models as smartphone agent","sources":["https://the-decoder.com/openmobile-2-proves-open-source-smartphone-agents-can-rival-proprietary-models/"],"verification":"reported","where":""},{"date":"2026-10-25","detail":"Released around Oct 25 with 110+ simulated apps (35 rebuilt popular apps including Taobao, JD, DiDi, Meituan, Slack, Google Drive) and two models built on Qwen3.5 9B and Qwen3.6 27B. The larger 27B model scored 59.6 on SPA-Bench on real Android devices, described as highest among open models, and improved from 67.1 to 79.9 on AndroidWorld and 35.9 to 50.4 on MobileWorld. With training data held constant, increasing app diversity pushed success on unseen apps from 60.7% to 67.5%.","from":"/ai/2026-10-27","headline":"OpenMobile-2 released Oct 25 with benchmarks showing open smartphone agents rivaling proprietary models","sources":["https://the-decoder.com/openmobile-2-proves-open-source-smartphone-agents-can-rival-proprietary-models/"],"verification":"researched","where":"global"},{"date":"2026-10-25","detail":"Built on three fine-tuned Qwen3.5 models (9B TempSeg, 27B Captioner, 27B/28B Rewarder), Flora is already used to train Nvidia GR00T models labeling up to 200,000 hours of video per day across 512 GB300 GPUs. Models are on Hugging Face as nvidia/Flora-Captioner, Flora-TempSeg and Flora-Rewarder. On benchmarks Flora leads on temporal segmentation (WGO-Bench F1@0.75 50.3 vs Astra 46.0) and failure recall (53.3% vs 0% on DenseRewardBench) but trails Astra on description accuracy. Cost claim of ~$0.72/hour vs ~$88.50 for GPT-6 Astra (~125x cheaper) is a projection assuming $10/GPU/hour, not measured production cost.","from":"/ai/2026-10-27","headline":"Nvidia open-sourced Flora on Oct 25 as open data engine for robot video labeling","sources":["https://the-decoder.com/nvidia-open-sources-flora-to-make-robot-video-labeling-125x-cheaper-than-gpt-6-astra/","https://research.nvidia.com/labs/gear/flora/","https://huggingface.co/collections/nvidia/flora"],"verification":"researched","where":"global"},{"date":"2026-10-25","detail":"As of Oct 25, LMArena and SWE-bench leaderboards not currently updating. Last measured records remain closed-model leads: LMArena text 1,532 (Gemini 4 Argon) vs open best 1,491, vision 1,325 vs 1,294, web dev 1,807 vs 1,665; APEX-Agents 82.4% vs open best 56.6%. No new independent scores for Muse Spark Open or Beam appeared in last three days.","from":"/ai/2026-10-25","headline":"No new independent benchmarks for Muse Spark Open or Beam in last three days; measured gap unchanged","sources":[],"verification":"reported","where":""},{"date":"2026-10-22","detail":"At GTC Berlin on Oct 22, Nvidia said European public-sector projects are building on open models: European Commission AI@EC via deepset Haystack, Poland eHealth Centre and Ministry of Labour agentic factory, UK Nemotron-Personas-UK, Germany Sinfonie and secunet SINA sovereign cloud, France CANUT Twake.ai, Greece KIEFER DGX SuperPOD.","from":"/ai/2026-10-24","headline":"Nvidia showcased European public-sector projects on open models at GTC Berlin Oct 22","sources":["https://blogs.nvidia.com/blog/european-public-sector-open-models/"],"verification":"researched","where":"Berlin, Germany"},{"date":"2026-10-22","detail":"At GTC Berlin on Oct 22, Nvidia said European financial services organizations are building proprietary intelligence on open models: Revolut PRAGMA on Blackwell Ultra (80M customers, 3x scale), Starling Intelligence on UK sovereign cloud, Lloyds ATUM on RTX PRO 6000, plus bunq, NatWest, Modulr and Santander with Kumo models.","from":"/ai/2026-10-24","headline":"Nvidia showcased European financial services building proprietary intelligence on open models at GTC Berlin Oct 22","sources":["https://blogs.nvidia.com/blog/european-financial-services-organizations-open-models-proprietary-intelligence/"],"verification":"researched","where":"Berlin, Germany"},{"date":"2026-10-21","detail":"CEO Jensen Huang keynoted GTC Berlin Oct 21 framing open models as sovereign intelligence and announcing DSX POD modular AI factories (deployment 6 months to 14 days), OpenShell Reference Architecture, and Agent Knowledge Foundation with Kumo models; portfolio expanded Oct 21 to include Nemotron Retriever, Kumo Structured/Engine, Cosmos 3.5 world models, Alpamayo 2 Edge, BioNeMo RFD4-Proteina and Ising, all with open weights/datasets.","from":"/ai/2026-10-24","headline":"Nvidia framed open models as sovereign intelligence at GTC Berlin keynote Oct 21 and expanded open portfolio","sources":["https://blogs.nvidia.com/blog/gtc-berlin-2026-keynote/","https://blogs.nvidia.com/blog/open-models-europe-industry/"],"verification":"researched","where":"Berlin, Germany"},{"date":"2026-10-21","detail":"On Oct 21 DeepSeek senior researcher Chen Deli wrote on social media that open vs closed source is a decisive battle between two futures for humanity and they will fight until total victory; Harness lead Cui Tianyi said they hope more businesses worldwide will deploy DeepSeek models and profit, per SCMP reporting on the DeepInfer release.","from":"/ai/2026-10-21","headline":"DeepSeek researchers framed open vs closed as decisive battle Oct 21","sources":["https://www.scmp.com/tech/big-tech/article/3371592/deepseek-open-sources-new-tool-slash-ai-costs-fight-against-closed-source-rivals"],"verification":"researched","where":"global"},{"date":"2026-10-20","detail":"DeepSeek open-sourced DeepInfer on Oct 20 under Apache 2.0 as its production inference framework powering V4.1-Flash. VLLM-based with FlashMLA, DeepGEMM and DeepEP kernels, it supports offline inference and runs on NVIDIA SM100 (Blackwell) GPUs and Huawei Ascend 950 NPUs, per GitHub and reporting.","from":"/ai/2026-10-21","headline":"DeepSeek open-sourced DeepInfer inference engine Oct 20","sources":["https://github.com/deepseek-ai/DeepInfer","https://the-decoder.com/deepseek-open-sources-the-inference-engine-that-powers-its-fastest-and-cheapest-ai-model/"],"verification":"researched","where":"global"},{"date":"2026-10-17","detail":"DeepSeek-V4.1-Flash-1017 shipped Oct 17 via website/app/API and as MIT-licensed open weights on Hugging Face/ModelScope, described as minor update to V4.1-Flash with RL with self-feedback. Reported 77.0% on DeepSWE v1.1 at ~$0.20/task vs predecessor 74.2% at ~$0.50/task, plus Search & Fetch APIs at $0.75 per 1,000 calls.","from":"/ai/2026-10-21","headline":"DeepSeek-V4.1-Flash-1017 shipped Oct 17 as open-weight cost breakthrough","sources":["https://the-decoder.com/deepseek-v4-1-flash-1017-beats-larger-rivals-in-agentic-tasks-while-once-again-massively-cutting-per-task-costs/"],"verification":"researched","where":"global"},{"date":"2026-10-16","detail":"Hugging Face blog/release published Oct 16 for Beam, a 501B-parameter MoE (23B active per token, 52 layers, 256 experts per MoE layer, 8 routed, 49 MoE layers + 3 dense), text-in text-out, 1M context, Apache 2.0 license, focused on coding and agentic use cases. Pretrained on 23.8T tokens with RL >100M rollouts on 10.5K NVIDIA GB300 GPUs over 4 weeks, 1M environments, 1.3B sandboxes. Checkpoints in BF16 (~1TB), FP8 and NVFP4.","from":"/ai/2026-10-19","headline":"Reflection's Beam published Oct 16 as 501B Apache-2.0 coding model","sources":["https://huggingface.co/blog/reflection-beam","https://reflection.ai/blog/introducing-beam"],"verification":"researched","where":"global"},{"date":"2026-10-16","detail":"Epoch AI Cyber ECI analysis published Oct 16 found Z.ai GLM-5.3 (released Aug 14, Cyber ECI 160.1) is top open-weight on cyber, above GPT-5.4 (Mar 5, 157.0) but trailing Claude Mythos Preview (Apr 6-7, 162-164) and GPT-5.5 (Apr 23, 162.2), placing open cyber frontier where closed frontier stood 6-7 months earlier; closed frontier as of Oct 8 is Claude Opus 5.5 and GPT-6 Astra (both 171). Anthropic Sep 29 analysis found GLM-5.3 most cyber-capable open-weight to date, lagging US frontier ~4 months on CAISI aggregate, with safeguards bypassable 64-100% and abliteration (2,200 GPU hours, ~$4,400) dropping refusal from >90% to ~3%.","from":"/ai/2026-10-19","headline":"Independent cyber benchmarks Oct 16 put open models 6-7 months behind closed frontier","sources":["https://epoch.ai/data-insights/open-closed-cyber-eci-gap","https://www.anthropic.com/research/glm-5-3-and-the-spread-of-advanced-cyber-capabilities"],"verification":"researched","where":"global"},{"date":"2026-10-15","detail":"Released Nemotron 3.5 Super VL (120B parameters) and Nemotron 3.5 Bolt 74B-A9B 24GB on Oct 15, with Bolt E4B 4GB scheduled Oct 23, covering data-center to edge. Super VL trained on ~26T tokens plus ~3.8M post-training samples with RL across dozens of environments; ships with open weights, open datasets and training recipes on Hugging Face and as NVIDIA NIM microservice. Bolt 74B uses 9B active MoE with 2-bit quantization to fit 24GB.","from":"/ai/2026-10-19","headline":"NVIDIA released Nemotron 3.5 Super VL 120B and Bolt 74B edge models Oct 15","sources":["https://blogs.nvidia.com/blog/nemotron-3-5-super-vl/"],"verification":"researched","where":"United States"},{"date":"2026-10-15","detail":"Released Oct 15 as open-weight frontier multimodal MoE with ~816B total parameters, ~42B active per token, 256 experts per MoE layer (8 routed), 62 layers, 1M-token context accepting text, images, audio and video and generating text. Weights under Muse Model License with limited commercial terms on Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden, Microsoft Foundry and hosted inference at Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, Together AI. Optimized to run locally on two Apple M5 Ultra Mac Studios (512GB unified memory each, 1.2 TB/s) via MLX holding ~1.5TB BF16 checkpoint (NVFP4 for Blackwell).","from":"/ai/2026-10-19","headline":"Meta released Muse Spark Open Oct 15 as 816B MoE with 1M context and desktop deployment","sources":["https://huggingface.co/blog/meta-muse-spark-open","https://research.meta.ai/blog/muse-spark-open","https://www.straitstimes.com/world/united-states/meta-unveils-new-open-ai-model-to-rival-chinese-competition"],"verification":"researched","where":"United States"}],"cite_as":"https://ohmega.monster/situations/thr_open_weights_race_muse_spark_open_nemotron_beam_","api_url":"https://ohmega.monster/api/v1/theaters/thr_open_weights_race_muse_spark_open_nemotron_beam_"},"meta":{"as_of":"2026-10-29T09:16:21.420746+00:00","source":"ohmega.dossier/1","version":"v1","cite_as":"https://ohmega.monster/situations/thr_open_weights_race_muse_spark_open_nemotron_beam_","license":"Ohmega claims, summaries, grades, Pulses and forecasts may be quoted with a link to cite_as. Source texts belong to their publishers; follow the source URLs. No warranty: readings are our assessment as of as_of, not guarantees.","note":"timeline[] and on_record[] items carry their own verification and source URLs."}}