Reasoning models, chain-of-thought, math performance, and test-time compute. — 90 releases covered on the show.
Microsoft launches MAI-Thinking-1, a 1T MoE trained from scratch
Microsoft AI used Build 2026 to launch seven MAI models, headlined by MAI-Thinking-1, a 1T total, 35B active MoE reasoning model trained from scratch on 33T tokens without distillation. The panel read the launch as Microsoft becoming a frontier model lab in its own right rather than only an OpenAI distribution channel.
NVIDIA releases Nemotron 3 Ultra, a 550B open-weight MoE for agents
NVIDIA dropped Nemotron 3 Ultra the day of the show, a 550B-parameter sparse MoE with 55B active parameters built for long-running agentic harnesses like OpenCode, Hermes, and OpenClaw. Chris Alexiuk joined to explain the hybrid Mamba/Transformer architecture and the unusually complete open release: weights, training data, recipes, a GenRM reward model, and an NVFP4 quantized checkpoint.
Anthropic ships Claude Opus 4.8 live mid-show
Anthropic released Claude Opus 4.8 during the episode, hitting 69.2% on SWE-bench Pro (up from 64.3% on 4.7 and ahead of GPT-5.5 at 58.6%), a new-best 57.9% on Humanity's Last Exam with tools, and 83.4% on OSWorld-Verified. It also shows a real long-context jump past the usual 200K cliff (85.9% GraphWalks BFS at 256K), with new thinking modes in the UI. Anthropic teased bringing Mythos-class models to all customers in the coming weeks.
Gemini 3.5 Flash launches at I/O as Google's agentic workhorse model
Google launched Gemini 3.5 Flash at I/O 2026 as a fast, determined workhorse model built for agentic loops rather than a budget-tier Flash like prior generations. It is rolling out across the Gemini app, Search AI Mode, the Gemini API, Google AI Studio, Antigravity and the Gemini Enterprise Agent Platform. Nisten noted unusual determinism in its behavior, and Logan Kilpatrick framed it as designed for the agentic era.
OpenAI model makes progress on 80-year-old Erdős planar unit distance problem
OpenAI announced that a general-purpose reasoning model made progress on the Erdős planar unit distance problem, challenging an 80-year-old mathematical belief. The panel called it the most important news of the week outside Google I/O, as a sign that frontier reasoning models are starting to contribute to genuinely open mathematics.
Mistral Medium 3.5: 128B dense flagship with 256K context
Mistral launched Medium 3.5, a 128B dense flagship model with 256K context and configurable reasoning, released with weights on Hugging Face. Alongside it Mistral shipped a Vibe coding agent.
GPT-5.5 and GPT-5.5 Pro drop live, SOTA across the board
OpenAI shipped GPT-5.5 and GPT-5.5 Pro mid-show, taking state of the art on Terminal-Bench 2 (82.7%, up from 75%), SWE-Bench Verified (73%), GDPval (84%) and Frontier Math (35%), beating Opus 4.7 and Gemini 3.1. It uses ~40% fewer tokens than 5.4, netting roughly 20% cheaper to run despite API pricing doubling to $5/$30 per million ($30/$180 for Pro). Peter Gostev called it the first model that genuinely sustains multi-hour long-running tasks, with one task running 8.5 hours straight; rollout was Codex-first, not yet in ChatGPT.
Claude Opus 4.7 drops live with 87.6% SWE-bench Verified and xhigh effort
Anthropic shipped Claude Opus 4.7 minutes before the show, scoring 87.6% on SWE-bench Verified and 64.3% on SWE-bench Pro, an 11-point jump over Opus 4.6 on the harder agentic coding eval. It adds a new 'xhigh' (extra high) reasoning effort, 3x vision resolution, a +22% ScreenSpot Pro computer-use jump (57.7% to 79.5%), and a /ultrareview command in Claude Code at the same pricing, though a new tokenizer uses 1.0-1.35x more tokens. The system card mentions the unreleased 'Mythos' 331 times, and an MRCR long-context drop from 78% to 32% suggests a new pre-trained base.
OpenAI's GPT-Image-2 leaks on LM Arena under three codenames
OpenAI's GPT-Image-2 posted the biggest single jump ever recorded on Arena, sitting 200+ ELO points above the previous top image model even on medium reasoning. The thinking/reasoning image model generates functioning QR codes, pixel-perfect infographics, 4K output, multi-image character consistency, and equirectangular 360-degree images that Peter Gostev stitched into a walkable street-view reconstruction of ancient Babylon. It even produces screenshots of IDEs containing SVG code that actually renders, enabling a new design-then-implement meta with Codex.
ARC-AGI-3 launches: humans score 100%, frontier models under 1%
ARC Prize launched ARC-AGI-3, an interactive agentic reasoning benchmark of turn-based puzzle games designed to test human-like generalization in novel abstract environments. Humans hit a 100% pass rate while top frontier models score under 1%, which the panel welcomed as a healthy reality check against AGI-is-here rhetoric and easy score inflation.
MiniMax M2.7: first self-evolving model hits 56% on SWE-Bench Pro
MiniMax dropped M2.7, billed as the first self-evolving model: it ran 100+ autonomous RL optimization loops and wrote its own agent scaffolding, built by one engineer over four days with zero lines of human code. It scores 56.22% on SWE-Bench Pro, within one point of Opus 4.6's 57.3%, and WolfBench shows it roughly matching Sonnet 4.6 on OpenClaw agent tasks. Not yet open weights, though rumors suggest a release is coming.
NVIDIA releases Nemotron 3 Super 120B with $26B open-source bet
NVIDIA launched Nemotron 3 Super, a 120B Hybrid Mamba-Transformer MoE model with 12B active parameters, a 1M-token context window, and 450 tok/s throughput. It shipped with BF16/FP8/NVFP4 weights, a base checkpoint, SFT and pre-training data, and the full training recipe, alongside a $26B 5-year open-source commitment. It is available on W&B Inference at $0.20/M input and $0.80/M output.
OpenAI drops GPT-5.4 Thinking and GPT-5.4 Pro live during the show
OpenAI released GPT-5.4 Thinking and GPT-5.4 Pro mid-show, a frontier general model that folds Codex-level coding into a unified reasoning model. It ships with a 1M token context window, a /fast mode, and mid-reasoning steering, posting 83.3% on ARC-AGI 2 (Pro) and roughly 75% on OS World computer use. The panel tested it live in Codex and called it a major general-model jump, while noting input pricing rose about 50% versus 5.2.
Agentica claims to solve all public ARC-AGI-3 tasks
Agentica published a claim of solving all public ARC-AGI-3 tasks, adding to the week's theme of benchmark saturation. The panel discussed it alongside METR and ARC-AGI-2 results as part of weighing signal versus noise in headline benchmark leaps.
Confluence Labs exits stealth with 97.9% SOTA on ARC-AGI-2
Confluence Labs emerged from stealth with a 97.9% state-of-the-art result on the ARC-AGI-2 benchmark, publishing code on GitHub. The panel read it as a major signal that ARC-AGI-2 is near saturation, part of a broader pattern of benchmarks getting solved faster than expected.
Gemini 3.1 Pro drops live with 44% HLE and 77% ARC-AGI at the same price
Google released Gemini 3.1 Pro minutes before the show, claiming 2.5x better abstract reasoning and improved coding and agentic capabilities at the same price point as its predecessor. It scores 44% on Humanity's Last Exam, 77% on ARC-AGI without a custom harness, and 68 on Terminal Bench, putting it at or near state of the art alongside Opus 4.6. In Nisten's live vibe-coding test it was blazingly fast but less polished than Opus 4.6 and Codex output.
Gemini 3 Deep Think scores 84% on ARC-AGI 2
Google dropped an upgraded Gemini 3 Deep Think mid-show, hitting 84% on ARC-AGI 2 — the biggest single jump in the benchmark's history, up from Opus 4.6's 68% set just one week earlier. It also scored 48.4% on Humanity's Last Exam without tools, taking state of the art on both.
Anthropic ships Claude Opus 4.6 with 1M context and agent teams
Anthropic dropped Opus 4.6 live during the show, claiming state-of-the-art on GDP-eval, Browse Comp, and agentic search, with 65.4% on Terminal Bench and 99% on TAU Bench MCP tool use. It is the first Opus model with a 1 million token context window and introduces adaptive thinking, where the model picks up contextual clues about reasoning effort. Pricing matches Opus 4.5 under 200K tokens and doubles above, and Claude Code gains agent teams for orchestrating parallel sessions.
Intern-S1-Pro: 1 trillion parameter open MoE for scientific reasoning
InternLM released Intern-S1-Pro, a 1 trillion parameter open-source MoE model targeting SOTA scientific reasoning across chemistry, biology, materials, and earth sciences. The panel noted it beats frontier models on science benchmarks, a massive compute investment for an open release.
StepFun Step 3.5 Flash: frontier reasoning claims at 11B active params
StepFun released Step 3.5 Flash, a 196B sparse MoE model with only 11B active parameters, claiming frontier-level reasoning while generating at 100-350 tokens per second. It continues the trend of sparse Chinese MoE models delivering high speed at low active parameter counts.
Google adds Agentic Vision to Gemini 3 Flash
Gemini 3 Flash gains agentic vision: a Think-Act-Observe loop that can zoom, crop, annotate, and plot images by generating and executing Python code in the backend. Available in the Gemini app, AI Studio, and Vertex AI.
Liquid AI's LFM2.5-1.2B-Thinking: on-device reasoning under 900MB
Liquid AI released LFM2.5-1.2B-Thinking, a 1.2B parameter reasoning model that runs entirely on-device with under 900MB of memory. Its hybrid architecture with gated convolutions delivers 239 tokens/sec on an AMD CPU and 82 tokens/sec on a mobile NPU, making it practical for edge devices, Raspberry Pi, and older iPhones.
Meituan's LongCat Flash Thinking: 560B MoE with 27B active, MIT licensed
Meituan released LongCat Flash Thinking, an open-source reasoning MoE with 560B total parameters and only 27B active, under an MIT license. It continued the run of large sparse Chinese open-weights models offering frontier-style reasoning at low active-parameter cost.
NVIDIA Alpha Mayo: open source reasoning self-driving models
NVIDIA announced Alpha Mayo at CES, a family of open source reasoning-based self-driving AI models. The models perform end-to-end autonomous driving with explicit reasoning steps, like identifying jaywalkers and stopping accordingly, demoed in a Mercedes-Benz.
Claude Opus 4 drops in Q2 — Ryan's pick for best model ever
Claude Opus 4 launched in Q2 and became Ryan Carson's pick as the best coding model he had used in over 700 days of daily LLM coding. It cemented Anthropic's lead in agentic coding through the middle of the year.
DeepSeek R1: the open reasoning model that crashed NVIDIA's stock
DeepSeek's open-weights reasoning model dropped January 23rd and matched OpenAI's o1 at roughly 50x cheaper pricing, with an alleged training cost of just $5.5M. It crashed NVIDIA stock 17% — a $560B single-day loss, the largest single-company monetary loss in history — and made Chinese AI a household topic. The crew named it the earthquake that shattered assumptions about who leads AI.
DeepSeek V3.1 Terminus lands amid September's relentless pace
DeepSeek resurfaced in September with V3.1 Terminus, another strong open-weights release that arrived just as the crew was barely keeping up with the weekly firehose. Nisten noted that missing a single week in this period left you completely lost.
Gemini 2.5 takes the #1 benchmark spot in March
Gemini 2.5 briefly claimed the top benchmark position in March, the moment Wolfram identified as the pivotal point where OpenAI stopped being the undisputed leader. It foreshadowed Google's full comeback later in the year.
Gemini 3 Flash delivers frontier intelligence at $0.50/1M input tokens
Google launched Gemini 3 Flash, offering frontier-tier capability at flash-tier pricing of $0.50 per million input tokens. It scores 78% on SWE-bench Verified, beating larger models on some agentic tasks, and supports tool-calling at scale with up to 100 simultaneous function calls.
Amazon announces Nova 2 family: Lite, Pro, Sonic, and Omni
Amazon rolled out the Nova 2 model suite spanning text, speech, and multimodal stacks with Lite, Pro, Sonic, and Omni variants. The launch came with major benchmark jumps over the first Nova generation and includes a fast, cost-effective reasoning model in Nova 2 Lite.
DeepSeek V3.2 and V3.2-Speciale post gold-medal reasoning under MIT license
DeepSeek released V3.2 and the reasoning-first V3.2-Speciale, a 685B-parameter MoE under MIT license. Speciale posted gold-medal-level olympiad results and 96% on AIME (versus GPT-5 High at 94%), with V3.2 hitting 73.1% on SWE-Bench Verified. Aggressive pricing around 28 cents per 1M tokens on OpenRouter pushes open models closer to top closed-model capability.
Gemini 3 Deep Think hits 45.1% on ARC-AGI-2 with parallel reasoning
Google shipped Deep Think, a high-cost parallel reasoning mode for Gemini 3 that scored 45.1% on ARC-AGI-2. The panel framed it as Google pressing its advantage in the frontier race, where product integration and latency now matter as much as raw benchmark IQ.
Anthropic launches Claude Opus 4.5, reclaiming the coding crown
Anthropic released Claude Opus 4.5, scoring 80.9% on SWE-bench Verified to top GPT-5.1 (77.9%) and Gemini 3 Pro (76.2%). It adds a new 'Effort' parameter for compute control, Tool Search to cut agent token overhead, and Programmatic Tool Calling where the model writes and executes code loops. Pricing dropped to $5/M input and $25/M output, roughly one-third the old Opus price.
DeepSeek Math V2: 685B open-weights model with IMO gold-level math
DeepSeek surfaced DeepSeek Math V2, a 685B-parameter Apache-2.0 model that reaches IMO gold-level math reasoning. It is the first open-weights math champion at this level, dropped quietly on HuggingFace during the week.
Prime Intellect releases INTELLECT-3, a 106B open MoE model
Prime Intellect released INTELLECT-3, a 106B-parameter mixture-of-experts model with 12B active parameters that scores 90% on AIME 2024/2025. The lab fully open-sourced the training stack alongside the weights, showing a small lab can train frontier-scale models.
Gemini 3 Pro launches with record ARC-AGI-2 scores
Google's new frontier multimodal model with a 1M-token context window and huge reasoning gains, scoring 31.11% on ARC-AGI-2 (45.14% with Deep Think mode) — roughly double the previous SOTA — plus 81% on MMLU-Pro and major coding improvements. Amp switched to it as their default model on launch day, the first time they have ever switched defaults. Also rolling out across Gmail, Calendar, and AI Mode in Google Search.
GPT-5.1 Pro: research-grade deep-thinking mode in ChatGPT
OpenAI also shipped GPT-5.1 Pro, a new research-grade ChatGPT mode that will happily think for minutes on a single query. It targets hard research-style questions where extended deliberation pays off, rounding out OpenAI's big week alongside Codex-Max.
Grok 4.1 briefly tops LM Arena with major post-training upgrade
xAI's Grok 4.1 shipped in November alongside GPT-5.1 and Claude Opus 4.5 in the year's most concentrated stretch of frontier releases. Yam highlighted the week-and-a-half window as emblematic of 2025's relentless acceleration.
Baidu open-sources ERNIE-4.5-VL-28B-A3B-Thinking visual reasoning model
Baidu released ERNIE-4.5-VL-28B-A3B-Thinking, an Apache 2.0 open-weights visual reasoning MoE with only 3B active parameters that claims to rival much larger models like GPT-5 High on vision tasks. It features image zooming, spatial grounding, and reasoning, with strong small-model performance attributed to GSPO training from the Qwen team.
OpenAI launches GPT-5.1 with a warmer, more personable voice
OpenAI shipped GPT-5.1, an update to its flagship model focused on a warmer tone and personality upgrades. The panel discussed how the friendlier default voice changes day-to-day ChatGPT use and what it signals for the frontier model race.
WeiboAI releases VibeThinker-1.5B open reasoning model
Weibo's AI team open-sourced VibeThinker-1.5B, a tiny reasoning model that reportedly outperforms much larger models like DeepSeek R1 on select reasoning benchmarks. Part of a week where small open-weights models from Chinese labs kept punching above their weight.
Moonshot AI releases Kimi K2 Thinking, an open 1T-param reasoning MoE
Moonshot AI released Kimi K2 Thinking, an open-source 1-trillion-parameter mixture-of-experts reasoning agent with 256K context and large-scale tool-calling capacity. The panel treated it as the open-source centerpiece of the week, focusing on its reasoning quality and coding utility rather than just benchmark screenshots, and as a sign open models keep closing the usability gap with frontier closed models.
OpenAI ships GPT-OSS-Safeguard, first open-weight safety reasoning models
OpenAI released GPT-OSS-Safeguard, its first open-weight safety reasoning models, built on the GPT-OSS family. The models let developers apply custom safety policies via reasoning rather than fixed classifiers, extending OpenAI's open-weights push into the trust-and-safety layer.
DeepSeek V3.1 Terminus refines agents and bilingual output
DeepSeek released V3.1 Terminus, an update to V3.1 with cleaner bilingual output, stronger agentic tool use, and cheaper long-context handling. The open weights are available on Hugging Face, continuing DeepSeek's cadence of iterative open releases.
xAI ships Grok 4 Fast with 2M context at a fraction of the cost
xAI released Grok 4 Fast, a cost-efficient model with a 2M token context window that unifies reasoning and non-reasoning behavior in one set of weights and prices far below Grok 4. The panel treated it as part of the larger competitive pressure cycle on price and speed among frontier labs.
Jeremy Berman and Eric Pang set new ARC-AGI SOTA using Grok-4
Independent researchers Jeremy Berman and Eric Pang published a new state-of-the-art result on ARC-AGI, built on Grok-4 with heavy test-time compute and iterative program synthesis. Berman joins the show to walk through the method, its limitations, and why iteration matters more than leaderboard narratives; the approach is documented in a detailed write-up.
Luma's Ray3: a 'reasoning' video model with native HDR
Luma AI launched Ray3, a video generation model it bills as a 'reasoning' video model, with native HDR output, a fast Draft Mode, and Hi-Fi mastering. It is available in Luma's Dream Machine and feeds the episode's closing theme of a next wave of video models.
Mistral updates its open reasoning model with Magistral-Small-2509
Mistral published Magistral-Small-2509, an updated checkpoint of its small open-weights reasoning model. The refresh keeps Mistral's open reasoning line current as the open-model competitive baseline moves quickly.
OpenAI adds thinking budgets to the ChatGPT app
OpenAI rolled out thinking budgets in the ChatGPT app, letting users control how much reasoning effort the model spends on a request. It is a small but notable product lever for tuning the cost-versus-quality tradeoff of reasoning models.
Nous Research releases Hermes 4 14B compact hybrid reasoning model
Nous Research launched Hermes 4 at 14B, a compact hybrid reasoning model with tool calling designed for both local and cloud use. It extends the Hermes 4 family down to a size practical for local deployment while keeping reasoning and tool-use capabilities, with a full tech report published on arXiv.
Microsoft's MAI-DxO hits 85.5% on NEJM diagnostic cases vs 20% for doctors
Microsoft AI published MAI-DxO, a medical diagnostic orchestration system that reached 85.5% accuracy on challenging NEJM-style cases compared to roughly 20% for practicing physicians. The result is framed as a systems win rather than a single-model win, suggesting orchestration may outperform individual models in high-stakes expert workflows.
Tencent ships Hunyuan-A13B: 80B MoE with only 13B active params
Tencent released Hunyuan-A13B-Instruct, an 80B-parameter MoE that activates only 13B parameters at inference while keeping a 256K context window. Built by the team with WizardLM lineage, it posts strong reasoning benchmarks and feels unusually practical for its class, though the panel flagged its license limits.
DeepSeek drops R1-0528, an updated open reasoning model with big gains
DeepSeek released R1-0528 out of nowhere, an update to their open-weights reasoning model with serious performance jumps: AIME 91, LiveCodeBench 73, and SWE-bench Verified 57.6. They also shipped an 8B distilled version based on Qwen3 that can run on a laptop, keeping it among the best open-weight models available.
Paper: models can learn to reason without external rewards
A mind-bending paper showing that reinforcement learning with internal or even random rewards can improve reasoning models. Intuitor matched or exceeded some GRPO results (the external-reward framework DeepSeek popularized with R1) when finetuning Qwen2.5 3B, questioning how much of RL's gains come from the reward signal itself.
AM-Thinking v1: 32B dense reasoning model beats bigger MoEs at math and code
A 32B dense open-weights reasoning LLM from a new Chinese team that takes on much larger mixture-of-experts models and comes out on top for math and code, hitting 85.3% on AIME 2024, 70.3% on LiveCodeBench v5, and 92.5% on Arena-Hard. It supports a /think reasoning toggle, ships with a permissive license, is tooled for vLLM, LM Studio, and Ollama, and runs at 25 tokens/sec on a single 80GB GPU with INT4 quantization. A multilingual RLHF pass and 128k context window are in the works.
ByteDance publishes Seed1.5-VL, a 20B vision-language thinking model
ByteDance's Seed team published the technical report for Seed1.5-VL, a 20B-parameter vision-language model with thinking capabilities. It was covered among the big-company releases of the week, with the tech report shared on GitHub.
Alibaba open-weights the full Qwen 3 family under Apache 2.0
Alibaba released the entire Qwen 3 stack: two MoE models (235B total/22B active and 30B/3B active) plus six dense siblings from 32B down to 0.6B, all Apache 2.0 with day-one support in LM Studio, Ollama, vLLM, MLX and llama.cpp. The headline feature is a runtime hybrid 'thinking' toggle (/think and /no_think) that trades latency for reasoning depth. Trained on ~36T tokens with 128K context and 119-language coverage, the 235B MoE rivals DeepSeek-R1, o1, o3-mini and Gemini 2.5 Pro on coding and math.
Microsoft ships Phi-4-reasoning and Phi-4-reasoning-plus (14B, MIT)
Microsoft fine-tuned the 14B Phi-4 on 1.4M curated chain-of-thought traces (SFT) and added a small RL stage (Plus variant) to create two MIT-licensed reasoning models. They punch far above their weight: Phi-4-reasoning-plus outperforms DeepSeek-R1-Distill-70B on AIME 25 (78% vs 51%) and sits within a few points of the full 671B DeepSeek-R1, while running on a single GPU with explicit <think> scaffolding.
Xiaomi enters open weights with MiMo-7B, MIT-licensed reasoning family
Xiaomi's first open-weights release is a 7B dense family (Base, SFT, RL, RL-Zero) trained from scratch on 25T tokens with a multi-token-prediction objective and rule-verifiable reinforcement learning. The RL variant matches OpenAI o1-mini on benchmark suites despite being far smaller, scoring 55.4% on AIME 2025 and 49.3% on LiveCodeBench v6, all under an MIT license with vLLM-ready weights.
Google launches Gemini 2.5 Flash with controllable thinking budgets
Google answered OpenAI's launch week with Gemini 2.5 Flash, a fast reasoning model that introduces controllable thinking budgets so developers can dial how much the model reasons per request. It is available through the Gemini API and developer platform.
OpenAI launches o3 and o4-mini, SOTA reasoning models with tool use
OpenAI shipped o3 and o4-mini in ChatGPT and the API, with o3 setting new SOTA records on Codeforces, SWE-bench, MMMU and more. For the first time the models can use tools (web search, Python, image generation) during the reasoning process, and they can think visually by cropping, zooming and rotating images. o3 scored $65k on the Freelancer eval versus o1's $28k, and o4-mini hits 99.5% on AIME with a Python interpreter.
Prime Intellect launches INTELLECT-2, a 32B globally-distributed RL run
Prime Intellect released INTELLECT-2, a 32B reasoning model trained with globally decentralized reinforcement learning, a follow-up to the INTELLECT-1 decentralized pretraining run covered on the show in December. The release includes open weights on Hugging Face, a tech report, and the PRIME-RL training code.
Z.ai (formerly chatGLM) releases the GLM-4-0414 open-source family
Z.ai, the rebranded Zhipu AI / chatGLM team, released the GLM-4-0414 family of open-source models. The drop includes base, reasoning and rumination variants published on Hugging Face and GitHub.
ByteDance publishes Seed-Thinking-v1.5 reasoning model tech report
ByteDance's Seed team published Seed-Thinking-v1.5, a new reasoning model announced via a technical report on GitHub. It was mentioned among the week's open-source LLM news, though weights were not released at the time.
Deep Cogito debuts Cogito v1 Preview models from 3B to 70B, beating DeepSeek 70B
New lab Deep Cogito released the Cogito v1 Preview family of open models ranging from 3B to 70B parameters, claiming SOTA results at each size and beating DeepSeek's 70B distill. The models are available on Hugging Face, giving local AI enthusiasts the small-to-mid sizes Llama 4 skipped.
Moonshot drops Kimi-VL and Kimi-VL-Thinking, tiny A3B open vision models
Moonshot AI released Kimi-VL and Kimi-VL-Thinking, compact vision-language models with only ~3B active parameters (A3B MoE). The thinking variant adds reasoning to a tiny VLM, and both are available openly on Hugging Face.
NVIDIA ships Nemotron Ultra, a 253B pruned and distilled Llama 3.1-405B
NVIDIA released Nemotron Ultra, a pruned and distilled finetune of Llama 3.1-405B at roughly half the parameters (253B). Its benchmarks even included Llama 4 comparisons, showing the older finetuned Llama beating the new models on AIME, GPQA and more. It supports 128K context and fits on a single 8xH100 node for inference.
DeepCoder-14B: open RL-finetuned coder beats DeepSeek R1 and o3-mini on coding
Together AI and Agentica (UC Berkeley Sky Computing Lab) released DeepCoder-14B-Preview, a reasoning model finetuned with RL that beats DeepSeek R1 and even o3-mini on several coding benchmarks. The project aims to democratize RL: the team open-sourced the model, the training dataset, the Weights & Biases logs, and the eval logs. Guest Michael Luo from Agentica joined the show to discuss the release.
Gemini 2.5 Pro scores 24.4% on USAMO olympiad math, crushing the field
New evaluation results published this week showed Gemini 2.5 Pro scoring 24.4% on the USA Math Olympiad (USAMO), problems so hard that most top models score under 5%. The result showcases a step change in frontier reasoning ability on competition mathematics.
Dream 7B: a diffusion language model challenger unveiled
Researchers unveiled Dream 7B, a diffusion-based language model that posts strong benchmark results, notably on planning-style tasks like Sudoku, possibly because parallel generation handles global constraints better than autoregression. It hints at viable alternative LLM architectures, but the weights were not yet released at show time, so results could not be independently verified.
ARC-AGI 2 benchmark revealed, thinking models score just 4%
The ARC Prize Foundation revealed ARC-AGI 2, the next iteration of the abstract reasoning benchmark. Base LLMs score 0% and even thinking models only reach about 4%, showing how far current frontier models remain from human-level fluid intelligence.
Google reclaims #1 with Gemini 2.5 Pro thinking model
Google dropped Gemini 2.5 Pro, a thinking model that took the #1 spot as the best all-around LLM available, with massive jumps on benchmarks like AIME (up nearly 20 points) and GPQA. It inherits native multimodality and a 1M token context window, maintaining high accuracy even at 120k+ tokens on needle-in-a-haystack tests, with surprisingly low latency (~13 seconds on hard reasoning questions vs 45+ for others). Tulsee Doshi, head of product for Gemini models, joined the show to give the inside scoop.
ByteDance releases DAPO, an RL method that beats GRPO
ByteDance published DAPO, a reinforcement learning method for LLM post-training presented as an improvement over GRPO. The paper ships with an open GitHub implementation, making the technique reproducible for the open-source RL community.
LG open sources EXAONE and EXAONE Deep 32B reasoning model
LG AI Research open sourced its EXAONE family, headlined by EXAONE Deep 32B, a thinking/reasoning model. The release puts a large Korean lab's reasoning model in open weights on Hugging Face, and Alex published a live reaction video to the launch.
NVIDIA drops Llama-Nemotron reasoning models plus training dataset
NVIDIA released the Llama-Nemotron family, including Super 49B and Nano 8B reasoning models, announced around GTC. Alongside the open weights, NVIDIA published the Llama-Nemotron post-training dataset, giving the community both the models and the data recipe behind them.
OpenAI makes o1-pro available via API at $600 per 1M output tokens
OpenAI exposed its o1-pro reasoning model through the API for the first time, priced at $600 per million output tokens. The show jokingly framed the pricing as 'for oligarchs', but it makes OpenAI's highest-compute reasoning tier programmatically accessible.
Nous Research releases DeepHermes 24B and 3B hybrid reasoning models
Nous Research released DeepHermes hybrid reasoners at 24B (Mistral-based) and 3B sizes, models that can toggle between standard chat responses and long chain-of-thought reasoning. The 24B preview is available on Hugging Face as part of the week's wave of open-source reasoning model releases.
Reka Flash 3: 21B open-source reasoning model under Apache 2.0
Reka AI open sourced Reka Flash 3, a 21B parameter reasoning model released under an Apache 2.0 license and trained with the REINFORCE Leave One-Out (RLOO) reinforcement learning technique. It excels at chat, coding, instruction following, and function calling, with Nisten calling it possibly one of the best ~20B models available.
Qwen releases QwQ-32B reasoning model that matches R1 on some evals
Alibaba's Qwen team released QwQ-32B, an open-weights reasoning model that matches DeepSeek R1 on several evals despite being roughly 20x smaller at 32B parameters. Qwen tech lead Junyang Lin joined the show to announce it, and the episode dubbed it Alibaba's 'R1 killer' for bringing strong reasoning to a size that runs on consumer hardware.
Anthropic releases Claude 3.7 Sonnet, a coding beast with immaculate vibes
Anthropic shipped its long-awaited model update, Claude 3.7 Sonnet, which the crew called a coding BEAST with 'immaculate' vibes. It was one of the week's two huge model drops alongside GPT-4.5 and became an instant favorite for AI coding workflows like those discussed in the Windsurf interview.
Perplexity releases R1-1776, a censorship-free DeepSeek R1 fine-tune
Perplexity open-sourced R1-1776, a fine-tuned version of DeepSeek R1 designed to remove Chinese government censorship on topics like Tiananmen Square and Taiwanese independence. They used human experts to identify around 300 sensitive topics and built a censorship classifier to train the bias out, claiming no significant impact on standard eval performance. The name 1776 is a nod to American independence.
xAI launches Grok 3, claiming SOTA benchmarks and a 1M token context window
xAI dropped Grok 3 on Monday evening, claiming state-of-the-art performance on several benchmarks and a 1 million token context window, with heavy emphasis on agents and future reasoners. The launch was messy, with a bug serving Grok 2 to some users and an eval-methodology spat with OpenAI over best-of-N scores, but vibes shifted positive, with co-hosts calling the base model the best coding model out. It is free for now, 'until their GPUs melt', with no API yet for independent evaluation.
Exa ships free DeepSeek R1 chat demo with web search
Exa integrated DeepSeek R1 into a free hosted chat demo that combines the reasoning model with Exa's web search. Mentioned in the tools section as a no-cost way to try R1 grounded with live search results.
Open Thoughts releases OpenThoughts-114k reasoning dataset
An open reasoning dataset with 114k examples released by the Open Thoughts project to fuel open replication of reasoning models like DeepSeek R1. It gives the open-source community high-quality chain-of-thought training data for distilling and fine-tuning reasoning LLMs.
Perplexity adds DeepSeek R1 as a Pro reasoning model option
Perplexity integrated DeepSeek R1 into its Pro search product, letting subscribers choose R1 as the reasoning model behind answers. It was one of several tools that raced to host R1 on Western infrastructure within days of the model's release.
Berkeley TinyZero and RAGEN replicate DeepSeek R1-Zero
Berkeley researchers released TinyZero and RAGEN, open replications of DeepSeek's R1-Zero reinforcement-learning recipe on small models. The projects showed that R1-style emergent reasoning behavior can be reproduced cheaply, with training runs logged publicly on Weights & Biases.
Humanity's Last Exam: a deliberately unsaturated frontier benchmark
Humanity's Last Exam (HLE) launched as a new, very hard benchmark designed to stay unsaturated as models max out MMLU and math evals. It crowdsourced expert-level questions to measure frontier model capability where existing benchmarks are at 98-99% saturation.
DeepSeek R1: MIT-licensed open source reasoning model rivals o1
DeepSeek released R1, a state-of-the-art open source reasoning model under a permissive MIT license. It matches or beats OpenAI's o1 on key reasoning benchmarks while being fully open weights, and DeepSeek also shipped a family of distilled smaller models. The show called this the hottest week open source AI has ever had.
Google ships updated Gemini Flash Thinking with 1M context
Google released an updated Gemini Flash Thinking model (01-21) with a 1 million token context window, built-in code execution, and improved evals over the previous Thinking release. It pushes Google's reasoning-model line forward in the same week DeepSeek R1 landed.
RAT: pipe DeepSeek R1 reasoning into other models
Guest Pietro Schirano released RAT (Retrieval Augmented Thinking), a technique and tool that extracts DeepSeek R1's reasoning traces and feeds them to a cheaper, faster model like GPT-3.5 Turbo for the final answer. It showcases the new pattern of mixing open reasoning traces with closed completion models.
Follow Reasoning & Math and everything else in AI — live every Thursday.
