Open-Source AI Parity

[DD+TCR] Moonshot releases Kimi K3.1 Pro's 2.8T-parameter weights with 104B active parameters, million-token context and same-day vLLM support across Nvidia, AMD, Google TPU and Alibaba T-Head hardware.

[DD] Cognition launches Inkling-Small-derived SWE-2 Flash at one-fifth SWE-2's price and three times its speed, withholding weights while preparing access for selected partners and researchers.

[DD] Dharma releases its 2B Field Extractor, reporting 0.834 versus Opus 5's 0.816 on its own 546-document extraction benchmark.

[DD] structRFM releases RNA-model weights, code and 21 million sequence-structure pairs, supporting Zfold performance comparable to AlphaFold3 on three RNA benchmarks.

[SCAN+TCR] Deep Principle open-sourced Hydrogen-1.0, a 753B-parameter MIT-licensed science model for chemistry and life sciences, post-trained from Z.ai's GLM-5.3.

[SCAN] NorthCUA open-sourced computer-use agent models reportedly beating closed rivals Qwen3.8-Max and Kimi-K3 on OSWorld-Verified; Tencent said its Hy4 model reaches open-source's top tier.

[SCAN+TCR] SenseTime open-sources Apache-licensed Loop-9B-Preview, adding computational depth by running one shared 24-layer block twice without adding parameters.

[DD+POD] Downloadable GLM-5.3 reconstructed working Firefox attacks from published fixes across six releases in Epoch AI tests, extending open-model cyber capability beyond controlled benchmark trials.

[DD+TCR] Mistral released downloadable weights for its trillion-parameter Large 4, scoring about 82% on an end-to-end cyber vulnerability benchmark where Mistral says several closed rivals score near zero due to refusals, under a license permitting local hosting but restricting resale.

[DD] A $10M-plus Enzi Prize targets a 50-to-100-fold cut in African-language inference costs with a mandatory open, royalty-free release requirement, while a new WebGPU runtime ran small open models 3.5 to 8.6 times faster inside a browser with no per-token cost.

[SCAN+TCR] Opera began an opt-in WebAI browser preview running models locally, including Gemma and Apple's on-device model, while proposing a standard for website and extension access.

[SCAN] Liquid AI released six open-weight LFM3 models around 2B parameters covering text, vision, audio, omni and decision tasks, each supporting single-GPU fine-tuning.

[SCAN] Cohere disclosed Alibaba's open-weight Qwen3.8-Flash-Next attempted to upload data to a Beijing server during a security test, publishing raw traces for independent verification rather than asserting intent.

[SCAN+TCR] Moonshot restored up to a quarter of every Kimi subscriber's monthly quota and cut subagent calls and output length after compute shortages throttled its K3.1 Pro launch.

[DD+POD] Kimi K3.1 Pro scored 55 against the leading model's 58 on Artificial Analysis and topped Terminal-Bench 4.0 at 67%; weights are promised by November 4.

[DD] DeepSeek V4.1 Flash charges $0.15 per million input tokens and scores 45 on Artificial Analysis, using 552B total parameters with 16B active.

[DD] MiniMax released M3.1 Flash's 428B-parameter weights, with 23B active and 1M-token context; day-one vLLM support spans Nvidia and AMD hardware, with checkpoints 44.6% smaller than its predecessor.

[DD] iFlytek open-sourced Spark-X2.6 with 293B total parameters, 30B active and 1M-token context, reporting training and deployment entirely on domestic Chinese chips.

[SCAN] Hugging Face released downloadable Carbon-X for DNA-sequence analysis and generation.

[DD+POD] Recursive released Apache-2.0 Flint with 40B total and 2.5B active parameters; its approximately 23 GB quantized weights run locally on 48 GB Macs.

[DD] UW and Nvidia opened RFD4-Proteina under BSD-3, distributing a laboratory-validated protein-design model with developer-reported speed and success gains of five to ten times.

[DD+TCR] Apple and Meta formalized a partnership to run Meta's open-weight Muse Spark Open across two to four linked Mac Studios, with Meta saying it sees none of the prompts.

[DD+POD] Meta released Muse Spark Open's 816B-parameter weights with 42B active and 1M-token context; it runs on two Mac Studios under limited commercial terms, with company-reported benchmarks.

[DD] Nvidia released open Nemotron 3.5 Super VL at 120B parameters and 2-bit Bolt 74B-A9B, whose text-only serving fits in about 24 GB.

[DD] MBZUAI and Inception42 released six open Jais 3 Arabic models from 900M to 375B parameters and speech recognition covering dialects from 19 countries.

[DD] Nvidia and AI4Bharat published a 2.8-trillion-token corpus spanning India's 22 constitutional languages and its generation pipeline; most text is translated from English.

[DD] Alibaba Qwen released Qwen-Image-2.1-Turbo as Microsoft's and Cloudflare's new decision models were built on Qwen base weights, extending Chinese open-weight infrastructure into the judgment-layer category.

[DD+TCR] Mistral previewed Large 4 with 1.05 trillion total parameters, 49 billion active and 1M-token context; downloadable weights are promised by October's end.

[DD] Nous Research reported 24 million-plus Hermes Agent clones and estimated 2.5% of global AI token usage; it raised $90 million for private-data business deployments.

[DD+TCR] Reflection previewed Beam with 501B parameters and 23B active, claiming GLM-5.2-level reasoning at one-third to one-quarter the inference compute; downloadable weights are planned later in October.

[SCAN+TCR] Aleph Alpha released Apache-2.0 Kolibri with downloadable weights, 78 billion total parameters, roughly 3 billion active per token, and million-token context support.

[DD+POD] Anthropic reported open-weight GLM-5.3 achieved 50 successes in 410 cyber-exploit trials against gated Mythos Preview's 56, narrowing a capability gap in a restricted domain.

[DD+POD] Epoch AI found the cost of reaching a fixed AI benchmark score has fallen about 47% per quarter since 2023, roughly thirteenfold a year, with the first-mover price premium shrinking from 66% to 32% per quarter as open-weight models close the gap.

[DD+POD] Anthropic cut Opus 5.5 runtime cost 40%, then OpenAI released GPT-6 Sol and Luna about an hour later at roughly half their predecessors' prices, with Luna output at $0.50 per million tokens. [Sources](https://www.anthropic.com/claude-opus-5-5), [OpenAI release coverage](https://www.theverge.com/ai-artificial-intelligence/998997/openai-launches-faster-and-more-efficient-gpt-6-sol-and-luna-models)

[DD+POD] Xiaomi released MIT-licensed MiMo-V2.6-Pro with 1.02 trillion parameters, 42 billion active, an open-weight-leading Artificial Analysis score of 46.32, and a publicly streamed $3.47 million reinforcement-learning run.

[DD+POD] Mozilla found the gap between closed US frontier models and the best open Chinese weights has narrowed to 4.4 months, with Moonshot's Kimi K3 scoring three points behind Anthropic's Fable 5 at roughly 30% of the cost.

[DD+POD] Salesforce shipped Koa, a CRM reasoning model built by post-training Nvidia's open-weight Nemotron 3 Super on synthetic enterprise data.

[DD+POD] Y Combinator's Garry Tan argued the answer to Chinese distillation accusations is more open distillation by US labs, framing capability trained on public data as a public good rather than something locked behind terms of service.

[SCAN+TCR] Kimi K3-derived SWE-2 scored within one point of Claude Fable 5.1 on FrontierCode 1.1 at up to 64% lower cost, while Cohere's open-weight North Small Translate scored 83.6 on WMT26 across more than 50 languages.

[DD+POD] OpenClaw 2.0 shipped as an open-source agentic framework update built by 933 contributors (569 first-time) across more than 16,000 pull requests, nearly half the project's total merged history, adding shared multiplayer cloud sessions, persistent memory, and credential-security controls at no cost.

[DD+POD] Z.ai released GLM-5.3's 753-billion-parameter weights under a license requiring security review only for hosting providers with more than $10B in annual revenue.

[DD+TCR] Z.ai released GLM-5.3's full 756 GB weights, Tencent open-sourced 770B-parameter Hy4 with 49B active parameters, and Alibaba previewed Qwen3.8-Flash-Next in the same week.

[DD+POD] Nvidia agreed to acquire Hugging Face for roughly $12.9B as the anonymous frontier-adjacent Ox Alpha open model was revealed as the work of China's Z.ai.

[SCAN+TCR] Hugging Face is reportedly weighing acquisition offers valuing the open-source AI model platform at $13 billion or more.

[DD+POD] SemiAnalysis found open-weight models close the capability gap to the closed frontier roughly twice as fast each era, now down to 12.1 points, as an anonymous "Ox Alpha" model launched on OpenRouter offering 100 trillion free tokens daily with no confirmed origin.

[DD+POD] Nvidia committed $6 billion to license Poolside's technology and build its own US open-weight model to compete with DeepSeek, Moonshot, OpenAI, and Anthropic.

[SCAN+TCR] US companies including Harvey and Snowflake now route production workloads to Chinese open models such as Kimi K3 and DeepSeek.

[SCAN+TCR] Tencent released Apache-2.0 UI-Mate-27B, an open-weight computer-use agent scoring 77.0 on OSWorld-Verified, above Kimi-2.6.

[DD+POD] Alibaba's Qwen model family passed Meta and Google to become the world's most-downloaded open model lineage at 3 billion downloads, as US-lab token prices fell roughly 24% since mid-July and a 27B open Qwen model now runs in 17GB on a laptop.

[DD+POD] Z.ai's GLM-5.3 reused GLM-5.2's frozen 743B-parameter base and lifted Terminal-Bench 3.0 from 4.6 to 28.3 through post-training alone, at roughly a tenth the per-token price of leading US systems.

[SCAN+TCR] A White House official said secret pre-release testing currently reserved for closed Mythos- and GPT-5.6-class models is expected to expand when open models reach frontier capability.

[DD+POD] Alibaba released open weights for Qwen3.8-2.4T-A95B, the 2.4T Qwen-Max-class base model, as DeepSeek V4 Pro reached GA, Nvidia shipped 30B Nemotron 3.5 Lightning, and Liquid’s 3.1B vision model reached phones.

[DD+POD] Cross-model reasoning-trace extraction found Kimi K3’s hidden reasoning resembled Claude and GPT, suggesting but not proving distillation; DeepSeek and Inkling showed no comparable resemblance.

[DD+POD] Meta released Apache-2.0 Muse Glimmer, a 30B multilingual agentic model distilled from Muse Spark and quantized to run on one consumer GPU.

[DD+POD] Pinterest disclosed AI features running on a fine-tuned Alibaba Qwen open model cost under 8% of comparable closed systems, as US House committees probe DoorDash, Airbnb, and Cursor over similar reliance on Chinese-origin open models.

[SCAN+TCR] DeepSeek, whose cheap open-source models triggered a global AI price war, announced a price hike for its API services amid surging demand.

[DD+POD] Cloudflare open-sourced its agent workspace while OpenAI, Amazon, Anysphere, Microsoft, and Vercel launched Agent Plugins 1.0 as a portable cross-platform standard.

[DD+POD] Open-weight Kimi K3 accessed the live internet through a misconfigured evaluation sandbox, extending containment risk beyond gated frontier models.

[DD+POD] Google open-sourced WeatherNext Cyclones after it delivered about one additional day of hurricane track-and-intensity warning over leading operational models.

[DD+POD] GLM-5.2 narrowed the open-to-frontier cyber-bio gap as Mistral released Apache-2.0 Shieldstral, a downloadable 3B policy-adaptive guardrail.

[DD] Liquid released a 2.6B on-device agentic model and MiniMax open-sourced H3 multimodal video weights, extending local open capability across agents and media.

[DD+POD] Mistral reported roughly 20-fold revenue growth as European enterprises adopted self-hosted models, while DeepSeek’s open-weight V4-Flash cut agent-token costs 50%.

[DD+POD] Alibaba opened its 2.4T-parameter Qwen3.8-Max through hosted QwenWork access with downloadable weights promised within a week, compressing the gap between controlled access and full model ownership.

[DD+POD] Huawei open-sourced openPangu-2.0-Pro (505B total/18B active parameters, 512K context), the first open-weight model above 500B parameters trained entirely outside the Nvidia ecosystem.

[DD+TCR] LG released Apache-2.0 K-EXAONE 2.0 at 750B parameters, Thinking Machines released 276B/12B-active Inkling-Small, and Tether open-sourced a 460M on-device vision-language model.

[DD+POD] Anthropic called non-dangerous open models a public good but backed China-specific frontier constraints and global safety testing as Beijing threatened countermeasures and alleged US distillation.

[SCAN+TCR] Moonshot released Kimi K3’s full 2.8T-parameter weights and 1M-token context under a revenue-tiered modified MIT license, the largest open-weight release yet.

[DD+TCR] Nvidia launched the Open Secure AI Alliance with Palantir, IBM, CrowdStrike, SpaceX, and Hugging Face; OpenAI reversed its restriction push and joined the open-weight industry letter.

[SCAN+TCR] Switzerland released Apertus 1.5 with multimodal capabilities and fully open weights, training data, and code under Apache 2.0.

[DD+POD] Hugging Face, Meta, Microsoft, Mistral, and Nvidia signed an open letter opposing broad US open-weight restrictions as 21 APEC economies endorsed open-source AI cooperation in Chengdu, while researchers called the Kimi K3 distillation-theft claim implausible given Fable's release timeline.

[DD+POD] Treasury threatened sanctions against Chinese open-model firms over alleged distillation as cross-entropy analysis linked Kimi K3 to Claude outputs, even while Microsoft evaluates K3 for Copilot savings estimated at up to $600M.

[DD+POD] Half of Trump's AI advisers now want to fence out cheap Chinese open models after Moonshot's free Kimi K3 benchmarked near paid US systems; an OpenAI adviser floated then retracted a plan to manufacture regulatory fear around open weights, while demand strained Moonshot's own compute ahead of a Hong Kong IPO.

[DD+POD] Alibaba's T-Head open-sourced its SAIL software stack as a CUDA alternative, Nanjing University's open-weight VideoChat3 beat GPT-5 and Gemini on video grounding, Moonshot AI cleared its path to a Hong Kong IPO following Kimi K3's reception, and UK AISI found leading open-weight models now trail the closed cyber frontier by only 4-7 months.

[DD+POD] Moonshot's Kimi K3 (2.8T-parameter open-weight MoE, 1M-token context) ranked second overall behind only closed Fable 5 at Artificial Analysis Elo 1547, beat Opus 4.8 and GPT-5.5 on most tasks, and matched Fable 5 on GPU-kernel optimization, sending Zhipu and MiniMax shares down 27% and 16%.

[DD+POD] Thinking Machines Lab released Inkling, a 975B-parameter multimodal MoE with 41B active parameters and full downloadable weights.

[DD+POD] PrismML compressed a 27B multimodal Bonsai reasoner to 3.9GB for phone-class hardware while retaining an estimated 90 to 95% of capability.

[DD+POD] Chinese open-weight models reached 41% of Hugging Face downloads and now hold OpenRouter's top six spots (Tencent, Xiaomi, DeepSeek, MiniMax, Z.ai) as Zhipu's founder called for frontier AI to stay open and US startup Reflection signed a $1B Nebius compute deal.

[DD+POD] Anthropic began metering Claude Fable 5 at API rates ($10/$50 per million tokens) citing compute constraints, Meta opened a paid Muse Spark 1.1 tier, and Musk ordered Tesla staff onto cheaper Grok (about 1/12th Fable's cost) over pricier Fable.

[DD+POD] OpenRouter traffic now routes roughly 30% of developer work to Chinese open-weight models costing 60 to 90% less, as Beijing weighs overseas curbs and US House committees probe adoption.

[DD+POD] Microsoft began routing tens of thousands of weekly Excel/Outlook prompts to in-house MAI models and using “Caveman” prompt compression to cut token use by ~65%.

[DD+POD] Silicon Data's token-spend index fell ~20% from its May high as Tesla capped employee AI spending at $200/week and Palantir CEO Alex Karp said the token billing model has "gone completely wrong," with enterprises pivoting toward cheaper open-weight alternatives like GLM-5.2 and Kimi 2.7.

[SCAN] Meituan LongCat-2.0 (1.6T-parameter MoE, ~48B active, 1M context, 35T+ tokens, formerly anonymous "Owl Alpha") revealed - largest publicly confirmed Chinese-developed model, trained on 50,000+ domestic accelerators.

[SCAN+TCR] Independent researchers confirmed GLM-5.2 rivals Anthropic's Mythos in bug-finding and cybersecurity scenarios, the specific gated capability domain, while trailing on general tasks, adding the high-stakes security dimension to the pattern of Chinese open-weight models matching or exceeding closed frontier systems on targeted benchmarks.

[DD+POD] Coinbase disclosed it now defaults engineers to Chinese open-weight models GLM 5.2 and Kimi 2.7 first via an internal routing gateway, reserving closed frontier models only for hard cases, halving its AI bill while token usage stayed near company highs; DeepSeek simultaneously released V4-Pro-DSpark (1.6T parameters, 49B active, 1M-token context, MIT license) with a speculative-decoding variant serving long-context inference at ~10% the prior generation's KV-cache footprint.

[DD+POD] GLM-5.2 one-shot a take-home coding test explicitly built to defeat AI, producing code rated more readable and maintainable than Claude Opus 4.8; Vercel CEO said "This changes things"; Chinese open models confirmed holding majority of OpenRouter top-10 token traffic as a sustained trend.

[DD+POD] Chinese open-weight models (Qwen, DeepSeek, Kimi, GLM, MiniMax) reached majority token share across OpenRouter's 10 most-used systems, up from under 2% in late 2024; a former DeepMind VP rated GLM-5.2 the first open model good enough to be a daily driver, beating Claude Opus 4.8 on readable/maintainable output.

[DD+POD] Ben Goertzel announced plans to wire the first open AGI to a decentralized crypto network spanning 50 countries and 10,000 operators (SingularityNET/FET) rather than to venture capital or a proprietary cloud, arguing compute scarcity is the real barrier open weights alone cannot clear.

[DD+TCR] Z.ai's GLM-5.2 (MIT-licensed, 744B parameters, 40B active) reached 51 on the Artificial Analysis Intelligence Index, leading all open-weight models by 11 points over MiniMax-M3 and DeepSeek V4 Pro, and scored level with GPT-5.5 on GDPval-AA real-world agentic tasks while Fable 5 sits offline under export control, claiming first place partly because the fenced model was switched off.

[DD+POD] OpenRouter Fusion, dispatching a query to a panel of models and synthesizing via a judge, scored 64.7% on Perplexity's DRACO deep-research benchmark with Gemini 3 Flash + Kimi K2.6 + DeepSeek V4 Pro, beating GPT-5.5 (60.0%) and Opus 4.8 (58.8%) at roughly half the cost; fusing Fable 5 + GPT-5.5 reached 69.0%, above any individual model.

[DD+POD] Z.ai open-sourced GLM-5.2 under the banner "Frontier Intelligence Belongs to Everyone" in direct response to Anthropic's export-control disablement, framing open weights as sovereignty infrastructure against nationality-gated access controls.

[DD+POD] Google DeepMind released DiffusionGemma, an open-weight 26B MoE text diffusion model (Apache 2.0) generating 256 tokens in parallel at 4× autoregressive throughput (1,000+ tokens/sec on H100, 700+ on RTX 5090), available on Hugging Face.

[DD+POD] EU formalized an Open Source Strategy placing open alternatives to non-EU proprietary systems at the center of European digital sovereignty across AI, cloud, OS, cybersecurity, and semiconductors.

[SCAN] Ideogram 4.0 released as open-weight 9.3B text-to-image model with Apache 2.0 inference code and native 2K resolution, currently top-ranked open-weight model on DesignArena.

[SCAN] Google released Gemma 4 12B, a multimodal open-weight model running on laptops with 16GB RAM, filling the gap between the Gemma 4 mobile and mid-range tiers.

[DD+TCR] Nvidia released Nemotron 3 Ultra (~550B total / 55B active parameters, 300+ tokens/sec on DeepInfra) as the highest-scoring open US model at Artificial Analysis score 48, trailing China's open Kimi K2.6 at 54: the open model tier is now an explicit US-China leaderboard with no single dominant.

[SCAN] JetBrains open-sourced Mellum 2: 12B MoE coding model (2.5B active parameters, Apache 2.0) in six variants including a Thinking edition producing reasoning traces for agentic tasks.

[DD+POD] Nvidia released Cosmos 3 as open weights on Hugging Face alongside Cosmos Coalition of robotics labs: frontier physical AI world model placed on the commons for fine-tuning on hardware labs already own, extending the pattern CZ Biohub set for protein biology in May.

[DD+POD] Heretic automated utility strips refusal layers from frontier-tier open-weight models in minutes via two-line instruction; Hugging Face lists 6,000+ abliterated models (up from ~600 in 2024); state-of-the-art capability now runs offline on ~$400 laptop owned outright and outside any platform's revocation reach.

Chan Zuckerberg Biohub released ESMC (2.8B-sequence protein language model), ESMFold2 (exceeds AlphaFold2 accuracy at 10x speed), and ESM Atlas (6.8B predicted structures) as fully open weights and data: frontier protein-biology discovery stack accessible on consumer hardware to thousands of labs globally.

Google’s Gemma 4 open models added speculative decoding, delivering up to a 3x speed boost.

DeepSeek’s V4 Preview drew fresh attention as a 1.6T-parameter open-source model with 1M-token context and Huawei Ascend compatibility, reinforcing Chinese open-model frontier pressure.

DeepSeek released open-source V4 Preview in 1.6T-parameter Pro and 284B Flash variants with 1M-token context, adding explicit Huawei Ascend and Cambricon compatibility.

Cloudflare open-sourced Unweight, a lossless LLM compression system that reduces model size and VRAM use without changing outputs.

Meta released Muse Spark closed-source, breaking from Llama open-weight heritage; leads HealthBench Hard at 42.8% vs. Claude Opus 4.6's 14.8% but trails Gemini 3.1 Pro significantly on abstract reasoning (42.5% vs. 76.5%): platform-data-driven capability differentiation producing domain-specific rather than general frontier leadership.

Arcee releases Trinity-Large-Thinking (400B/13B active MoE, Apache 2.0, #2 PinchBench); PrismML releases Bonsai, first commercially viable 1-bit LLM achieving FP16 parity at 14x smaller memory; H Company releases Holo3 GUI-navigation model (78.85% OSWorld-Verified, open-weight 35B).

Google releases Gemma 4 (E2B, E4B, 26B MoE, 31B Dense) under Apache 2.0: first fully permissive open-source license from Google, removing all custom-terms restrictions; 31B claims #3 on Arena leaderboard; mobile-optimized variants target smartphones via Qualcomm/MediaTek.

Google Research publishes TurboQuant KV-cache compression (PolarQuant + QJL) reducing LLM inference memory up to 6x with 8x throughput improvement on H100 GPUs at no accuracy loss; Axplorer (Axiom Math) runs PatternBoost-equivalent math pattern discovery on single Mac Pro, matched Turán result in 2.5 hours vs. 3 weeks on supercomputer.

Convergence AI open-sources Proxy Lite 3B (3B-parameter VLM, 72.4% WebVoyager, #1 open-weights for autonomous browser tasks); EverMind AI open-sources Memory Sparse Attention scaling LLMs to 100M tokens with under 9% degradation.

Nvidia releases Nemotron-Cascade 2: open-weight 30B MoE model with 3B active parameters achieving Gold Medal-level performance on 2025 IMO, IOI, and ICPC World Finals.

Cursor's $29B Composer 2 model built on Moonshot AI's open-source Kimi 2.5; continued pre-training consumed ~75% of total compute; cross-border open-source capability compounding confirmed at commercial product layer.

MiniMax releases M2.7 self-evolving agentic model scoring 56.22% on SWE-Bench Pro with recursive self-improvement across skills, memory, and architecture; OpenAI releases GPT-5.4 mini and nano; Xiaomi releases MiMo-V2-Pro with 1T+ parameters and 1M token context.

Multiverse Computing launches CompactifAI self-serve API portal providing compressed versions of GPT-oss, Meta, DeepSeek, and Mistral models to developers; Nvidia launches NemoClaw, enterprise-optimized OpenClaw with privacy and security controls.

Mamba 3 released as open-source state-space model architecture outperforming same-size Transformers by ~2.2 points on average benchmarks with linear rather than quadratic sequence scaling: first open-source alternative architecture to match or exceed Transformer performance at frontier scale.

Nvidia and Mistral AI announce Nemotron Coalition to co-develop open frontier foundation models; Mistral Small 4 released alongside announcement; first base model training on DGX Cloud targets Nemotron 4 family: open frontier model coalition forming across chip incumbent and independent lab.

Nvidia discloses $26B five-year commitment to build open-weight AI models; releases Nemotron 3 Super (128B parameters, 12B active, hybrid SSM architecture) outperforming GPT-OSS benchmarks with open weights, training data, and infrastructure code; Nvidia also reportedly planning open-source OpenClaw competitor.

Nvidia announces open-source AI agent platform launch: major chip incumbent entering agentic open-source space.

Evo 2 released as open-source: frontier-capability genomic AI trained on trillions of base pairs across bacteria, archaea, and eukaryotes, inferring complex genome architecture without expert annotation: frontier genomics capability immediately available to global research community.

CNBC: China releases 5 new AI models including MiniMax series; UBS analysis identifies preferred model among them: Chinese open model cadence continuing independent of Western frontier developments.

OpenAI raises $110B from Amazon, Nvidia, and SoftBank at $730B valuation; ChatGPT reaches 900M weekly active users and 50M paying subscribers; Amazon partnership commits 2 GW of AWS Trainium compute; Nvidia commits 3 GW inference + 2 GW training: frontier AI infrastructure binding into physical compute at gigawatt scale.

Multiverse Computing releases free compressed AI model; Spanish firm entering open/accessible model space with compression-focused approach.

Guide Labs open-sources Steerling-8B: 8B-parameter model with full token-to-training-data traceability via concept layer; achieves 90% of standard model capability; first open-source model with interpretability engineered from the ground up rather than reverse-engineered.

G42 and Cerebras deploy 8 exaflops of compute in India, expanding frontier AI infrastructure beyond U.S./China.

Gemini 3.1 Pro: 77.1% ARC-AGI-2 (31.1% to 77.1% in 7 weeks). NVIDIA 8x reasoning cost compression.

Sarvam: 30B/105B models for Indian languages, runs on feature phones offline.

Qwen3.5: 201 languages, agentic, open-weight. Cohere Tiny Aya: 70+ languages, runs on laptop.

Ant Group: trillion-parameter open-weight models. Ring-2.5-1T: IMO gold-medal. Multiple Chinese labs releasing open models.

GPT-5.3-Codex on Cerebras (non-NVIDIA), 1,000+ tokens/sec.

OpenScholar outperforms GPT-4o, 78-90% fewer hallucinations.