Architecture innovations: MoE, attention variants, SSMs, diffusion LMs, and long context. — 69 releases covered on the show.
Inception Mercury 3: a diffusion LLM claiming 1,000+ tokens/sec
Inception released Mercury 3, a diffusion language model it says runs at more than 1,000 tokens per second with a 1M-token context. These are vendor numbers.
Meta releases Muse Spark Open: 816B open weights, 42B active, 1M context
Meta dropped Muse Spark Open during the last minutes of the show: 816B total parameters, 42B active, 1M context, multimodal in and text out, and Meta calls it the most capable US open-weight model to date. On Terminal-Bench 2.1 it scores 76.4 versus 83.9 for GLM 5.3 and 85 for Kimi K3. Nisten read the Muse Model License live: products must show 'built with Muse Spark Open' and there is a $20M annual revenue cap.
TypeSafe AI launches Jev, the first public non-LLM System 1 decision model
TypeSafe AI, the stealth lab led by RLHF and ChatGPT co-creator Diogo Almeida, launched Jev — the first public System 1 model, a new class of AI that returns calibrated probabilities instead of generating text. Developers define questions with three primitives (choice, score, null) in natural language and get typed, machine-readable probability outputs back in 70-500ms with a 32K context window, trained with what TypeSafe calls RLCD (reinforcement learning for calibrated decisions). Pricing is $42 per billion input tokens with free output tokens, and TypeSafe cites 133x faster and 444x cheaper than competitive-level LLMs on tested decision tasks. Access is via waitlist.
DeepSeek V4.1 Flash: 552B MoE, 8B/16B active, KV cache 400x smaller than V1, MIT
DeepSeek's V4.1 Flash is a 552B-parameter multimodal MoE that activates only 8B parameters on prefill and 16B on decode, with a 1M-token context window, trained from scratch on 45 trillion multimodal tokens and released under MIT. It returns to an encoder-decoder architecture at half-trillion scale and cuts the KV cache to about 890 bytes per token, down from 389,000 bytes in the first DeepSeek release (over 400x smaller), which is why it is so cheap to serve. DeepSeek's own evals put it at 90.6 on Terminal-Bench 2.1 and 74.2 on DeepSWE 1.1, above Opus 5 and GPT-5.6 Sol, and second behind GPT-6 Astra on an Open Design leaderboard at about two cents per task. TokenJuice.ai serves it free in the US for a limited time in exchange for training data.
Qwen3.8-Flash-Next previews the Qwen4 architecture in open weights
Alibaba released open weights for Qwen3.8-Flash-Next: 125B parameters plus a 51B N-gram embedding table with only 6B active, trained at roughly 1/9 the cost of Qwen3.7-Plus while self-reporting DeepSWE 58.7 and SWE-bench Pro 62.5. It previews the Qwen4 architecture: Qwen Sparse Attention (QSA) replaces full attention layers, and the low-bandwidth N-gram embedding table can be offloaded to slow RAM or even an SSD.
Meituan's LongCat-Flash-Lite-Sparse: 1M native context at 3B active parameters, MIT licensed
The 'DoorDash releases a model' moment: 69B total parameters with ~3B active per token, native 1M-token context (up from 256K dense), MIT licensed. LongCat Sparse Attention lifts SWE-Bench Verified from 54.4 to 68.2 and SWE-Bench Multilingual by 21 points over the dense twin, building on DeepSeek Sparse Attention while removing its O(L²) scoring overhead. The paper reports the architecture scaling to 560B-A27B.
Moonshot releases Kimi K3's full open-weight checkpoints — 2.8T parameters, the largest open model ever
Two weeks after the API launch, Moonshot published Kimi K3's full checkpoints, model code, and technical report: 2.8T total parameters with 104B active (16 of 896 experts), native vision, a 1M-token context window, and roughly 1.56TB of MXFP4 weights. The report details KDA linear attention, attention residuals, NoPE, and a claimed 2.5x scaling-efficiency jump over K2. On the show, Elie Bakouch called it public building blocks scaled superbly, and Baseten's Philip Kiely described serving it day-zero on eight GB300s. The custom license requires branding above 100M MAU or $20M monthly revenue and a signed agreement for model-as-a-service providers — every provider lists the identical $3/$15 price.
Google quietly patches Gemma 4 with Flash Attention 4 and tool-calling fixes — no version bump
Google shipped a stealth update to the Gemma 4 family: Flash Attention 4 support on Hopper-class GPUs (a reported 25-70% prefill throughput speedup), tool-calling bug fixes, reduced model 'laziness,' and configurable vision resolution. The ThursdAI panel criticized shipping new weights under the same Gemma 4 name with no version bump, leaving users unsure which checkpoint they're actually running.
Moonshot AI open-sources Kimi K2.7 Code for agentic coding
Moonshot AI open-sourced Kimi K2.7 Code, a trillion-parameter MoE coding model with benchmark jumps over K2.6 and fewer reasoning tokens. On the show it landed as the second half of the open-source coding wave beside GLM-5.2.
Z.ai releases GLM-5.2, a 753B open MoE with 1M context
Z.ai released GLM-5.2 as a major open-source coding and agentic model: a 753B-parameter MoE, MIT-licensed, with a one-million-token context window. The episode treated it as the open-source model that arrived exactly as Fable access disappeared, with strong coding and agentic performance close to the frontier.
MiniMax announces M3 coding/agentic model with 1M context
MiniMax announced M3, a natively multimodal coding and agentic model with a one-million-token sparse attention context claim and open weights promised soon. Reported numbers include 59 on SWE-bench Pro, and the panel noted MiniMax already has a following for cheap agentic tool calling even as pure coding quality is debated.
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.
Cohere releases Command A+, a 218B Apache 2.0 MoE with 25B active params
Cohere released Command A+, a 218B-parameter mixture-of-experts model with 25B active parameters, shipping open weights under Apache 2.0. It was the week's headline open-source release, available on Hugging Face in both W4A4 quantized and BF16 variants.
Nous Research publishes Lighthouse Attention for fast long-context pretraining
Nous Research released Lighthouse Attention, a sparse attention method for long-context pretraining that delivers major speedups. The release includes a blog post, an arXiv paper and an open-source GitHub implementation.
Krea 2: Krea's first from-scratch foundation image model
Krea released Krea 2, its first foundation image model trained from scratch, built over six to seven months by nearly half the company. It focuses on aesthetic diversity, style control with up to 4 reference images, and moodboard-driven workflows, generating images in roughly 15 seconds. Co-founder and CEO Victor Perez joined the show to walk through it.
DeepSeek V4: 1.6T MoE with CSA+HCA attention and 1M context
DeepSeek released the V4 paper and models (V4-Pro and V4-Flash on Hugging Face), a 1.6T-parameter MoE featuring CSA+HCA attention that fits 1M tokens of context in just 5.7GB of KV cache. It is possibly the first frontier model trained across multiple datacenters, and DeepSeek is offering API tokens at an 80% discount on already much cheaper pricing.
NVIDIA Nemotron 3 Nano Omni: hybrid Transformer-Mamba MoE
NVIDIA released Nemotron 3 Nano Omni, a 30B-total/3B-active hybrid Transformer-Mamba MoE with 256K context. It delivers 9x throughput on consumer hardware.
SenseTime open-sources SenseNova U1 unified multimodal MoE
SenseTime open-sourced SenseNova U1, a unified multimodal MoE model with 8B total and 3B active parameters that handles understanding and generation with no separate encoder or VAE. The architecture builds on a paper the team presented at ICLR last year.
Gemma 4 21B REAP: 20% expert-pruned Gemma 4 26B MoE
Community researcher 0xSero released Gemma 4 21B-A4B REAP, a 20% expert-pruned version of the Gemma 4 26B MoE created using Cerebras' REAP pruning technique. It shrinks the model for cheaper local inference while preserving most of its quality.
Qwen 3.6-35B-A3B: Apache 2.0 MoE with 3B active hits 73.4% SWE-Verified
Alibaba Qwen open-sourced Qwen 3.6-35B-A3B under Apache 2.0 the same morning Opus 4.7 dropped: a 35B MoE with only 3B active parameters that scores 73.4% on SWE-bench Verified, rivaling models 10x its size. It is natively multimodal with 262K context extensible to 1M, and the crew called it the strongest mid-size LLM on nearly all benchmarks, putting to rest doubts about Qwen's open-source commitment after Junyang Ling's departure.
Baidu ERNIE-Image: 8B DiT ranks #1 on GenEval among open models
Baidu released ERNIE-Image, an 8B diffusion transformer that ranks #1 on GenEval among open models and features precise multilingual text rendering. It is part of this week's wave of Chinese open releases in image and 3D generation.
Alibaba ships Qwen3.6-Plus with near-Opus agentic coding and 1M context
Alibaba released Qwen3.6-Plus, an API model with agentic coding performance near Opus 4.5 and a 1M-token context window. The panel noted continued strong momentum for the Qwen family in practical coding and agent workloads.
Anthropic makes Opus 4.6 1M context the default in Claude Code, same price
Anthropic made 1M token context the default for Opus 4.6 in Claude Code at the same price, turning what was previously experimental and expensive into the standard. MRCR benchmark performance holds at 93% at 256K and 76% at 1M. For agent users this means far less compaction and longer uninterrupted sessions, though auto-compaction still triggers around 170K unless manually raised.
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.
Mistral Small 4: 119B MoE with 6B active unifies vision, coding, reasoning
Mistral returned to open source with Small 4, a 119B-parameter MoE with 128 experts and only 6B active per token, released under Apache 2.0. It unifies the previous Pixtral (vision), Devstral (coding), and Magistral (reasoning) lines into one model and can fit on a single H100 when compressed. Early WolfBench results are sobering at ~17% on OpenClaw agent tasks, roughly on par with similarly sized Nemotron.
Mamba-3 lands with three SSM innovations for inference-first linear models
Mamba-3 dropped with three SSM-centric innovations: trapezoidal discretization, complex-valued states, and a MIMO formulation aimed at inference-first linear models. It extends the state-space model line that underpins the growing wave of hybrid SSM architectures for long-context and agentic workloads.
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.
Black Forest Labs introduces Self-Flow
Black Forest Labs published Self-Flow, new research from the FLUX makers in the AI art and diffusion space. It was included in the week's AI Art & Diffusion roundup.
Google launches Gemini 3.1 Flash-Lite with 1M context at 360 tok/s
Google launched Gemini 3.1 Flash-Lite, a fast and cheap model with 1M token context aimed at the instant/fast tier, running around 360 tokens per second. The panel flagged a material pricing jump versus the prior Flash-Lite generation but saw it as well suited for judge, guardrail, and orchestration workloads in agent systems.
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.
Qwen 3.5 lands: 35B/3B-active Medium outperforms the old 235B flagship
Alibaba released the Qwen 3.5 family of open-weight models, headlined by Qwen3.5-35B-A3B, a 35B model with only 3B active parameters that outperforms their previous 235B flagship. Variants include a 122B-A10B and a dense 27B, with the panel highlighting the hybrid state-space (Mamba-layer) architecture and strong practical coding and agent performance at a tiny active-parameter footprint.
Liquid AI releases LFM2-24B-A2B, a laptop-friendly 24B MoE
Liquid AI released LFM2-24B-A2B, a 24B mixture-of-experts model with only 2.3B active parameters that runs on consumer laptops. The panel highlighted its speed and surprisingly strong non-coding reasoning, reinforcing the trend of efficient low-active-parameter open models for local use.
Alibaba opens Qwen 3.5: 397B-param multimodal MoE with only 17B active
Alibaba released Qwen3.5-397B-A17B, billed as the first open-weight native multimodal MoE model, with 397B total parameters, just 17B active, 512 experts, and 262K native context extendable to 1M. It delivers 8.6-19x faster inference than Qwen3-Max and continues Qwen's strength in multilingual and medical tasks, scoring 52.5% on Terminal Bench, third place among open-source models. Nisten found coding still trails GLM-5.
Anthropic ships Claude Sonnet 4.6 with 79.6% SWE-Bench and 1M context
Anthropic launched Claude Sonnet 4.6, its most capable Sonnet ever, scoring 79.6% on SWE-Bench Verified, nearly matching Opus 4.6 at Sonnet pricing of $3/$15 per million tokens. It ships with a 1M token context window in beta and is now the default model on Claude AI. In blind Claude Code testing, users preferred Sonnet 4.6 over the previous Opus 4.5 59% of the time, and it beats the previous Gemini 3 Pro on most benchmarks.
Arcee AI ships Trinity Large: 400B MOE trained in 33 days for $20M
Arcee AI's Trinity Large is a 400B-parameter MOE with 13B active parameters, trained on 17T tokens across 2000 B300 GPUs in 33 days for $20M. It has 512K native context (twice Kimi K2.5), is free on OpenRouter until February 2026, and the panel called it the largest Western open-source lab model.
Sakana AI's RePo lets LLMs dynamically reorganize their context
Sakana AI introduced RePo, a research technique that lets language models dynamically reorganize their context for better attention. The paper proposes a new way to manage what a model focuses on, aimed at improving performance on long-context tasks.
Upstage Solar Open 100B: 102B MoE trained on 19.7T tokens
Upstage released Solar Open 100B, a 102B parameter MoE model with only 12B active parameters per token (129 experts, top-8 activation), trained on 19.7 trillion tokens including 4.5T synthetic via a 'data factory' approach. It outperforms GLM 4.5 Air on many benchmarks, features the SNAP PO reinforcement learning technique with a 50% training speedup, and delivers best-in-class Korean language performance.
MiniMax-01: open model with a 4M token context window
MiniMax (Hailuo) released MiniMax-01 in January with a 4 million token context window, by far the largest context of any open-weights model at the time. It was an early sign of the Chinese-lab open source dominance that defined 2025.
Allen AI's BOLMO reaches byte-level parity with tokenized models
Allen AI released BOLMO, described as the first byte-level language model to reach parity with regular tokenization-based models. The panel framed it as a research breakthrough that could eventually remove tokenizers from the LLM stack.
NVIDIA ships Nemotron 3 Nano, a 30B hybrid Mamba-MoE with full recipes
NVIDIA released Nemotron 3 Nano, a 30B-parameter hybrid Mamba-MoE model with only 3B active parameters for efficient inference. The panel called it the most consequential open release of the week because NVIDIA shipped not just weights but technical reports, training recipes, and details on the 25T-token training data.
Arcee Trinity launches US-trained open MoE family
Arcee AI introduced Trinity, a family of US-trained open mixture-of-experts models built from scratch, starting with Trinity-Mini and Trinity-Nano-Preview. CTO Lukas Atkins joined the show to discuss the training approach and previewed Trinity-Large for January 2026. The release positions Arcee as a domestic alternative in an open-weights field dominated by Chinese labs.
Tongyi's Z-Image Turbo brings sub-second open image generation
Alibaba's Tongyi lab released Z-Image Turbo, a 6B-parameter open image generation model that produces images in under a second. It pushes open-source image generation toward real-time speeds at a fraction of the size of competing models.
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.
Tencent releases HunyuanVideo 1.5, a lightweight open video model
Tencent released HunyuanVideo 1.5, a lightweight DiT-based open-source video generation model. It brings capable video generation to a smaller footprint, continuing the trend of open video models closing the gap with closed offerings.
Qwen Image Edit gains Multi-Angle LoRA for camera control
A Multi-Angle LoRA for Qwen Image Edit landed, enabling camera-control style edits that re-render a scene from new angles. Available as a Hugging Face space and on fal, it shows the fast-moving open ecosystem building on Qwen's image editing models.
NVIDIA releases ChronoEdit-14B Upscaler LoRA
NVIDIA released an Upscaler LoRA for its ChronoEdit-14B image editing model, available on Hugging Face with Diffusers pipeline support. It adds high-quality upscaling to the ChronoEdit physics-aware editing stack.
Grok 4 Fast expands to a 2 million token context window
xAI's Grok 4 Fast now supports a 2 million token context window, one of the largest of any frontier model. The crew called the jump 'crazy' and discussed what such long context unlocks for agentic and document-heavy workloads.
Ming-flash-omni Preview: sparse MoE omni-modal open model
Ant Group's InclusionAI team released Ming-flash-omni Preview, a sparse mixture-of-experts omni-modal model on Hugging Face. It handles multiple input and output modalities in a single open-weights model, adding to the wave of Chinese open omni-modal releases.
Kimi Linear: 48B open model with linear attention and 1M context
Moonshot AI released Kimi Linear, a 48B parameter (A3B active) instruct model that uses linear attention to reach a 1M token context window. It is an open-weights bet on efficient long-context architectures from the Kimi team.
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.
Moondream 3 Preview: 9B MoE VLM with 2B active parameters
Moondream released a preview of Moondream 3, a 9B mixture-of-experts vision-language model with only 2B active parameters. It targets frontier-level visual reasoning at small-model cost, continuing Moondream's run of efficient open vision models.
Hunyuan SRPO: preference optimization that supercharges diffusion models
Tencent Hunyuan published SRPO (Semantic Relative Preference Optimization), a post-training technique that significantly improves the output quality of diffusion image models. The team released weights on Hugging Face along with a project page and striking before/after comparisons.
Huawei's Pangu Pro MoE: 72B model trained entirely on Ascend NPUs
Huawei released Pangu Pro, a 72B-parameter MoE trained on its own Ascend NPUs rather than Nvidia or AMD hardware, hitting 1,528 tokens/sec and pretrained on 13T tokens. The panel framed it as the geopolitical open-model story of the week, showing how far Chinese compute stacks have advanced under sanctions.
Mystery 1M-context model 'Cypher Alpha' appears free on OpenRouter
A stealth model called Cypher Alpha showed up on OpenRouter with a free 1M-token context window, with the panel speculating it could be Amazon Titan. Alex used it as an example of how model releases increasingly arrive as anonymous market probes rather than tidy launches.
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.
Alibaba's Wan 2.1: open-source diffusion-transformer text-to-video suite
Alibaba, the team behind the Qwen LLMs, released Wan 2.1, a full stack of open-source diffusion-transformer text-to-video foundation models. Amid the show's discussion of video-model fatigue, this was called out as a release that cuts through the noise, with weights on Hugging Face and code 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.
Sand AI surprises with MAGI-1, a 24B streaming autoregressive video model
Sand AI released MAGI-1, a 24B autoregressive diffusion model for long-form, streaming video generation with remarkable character consistency, often the Achilles' heel of AI video. It predicts video in 24-frame chunks with causal attention between them, enabling real-time streaming generation where compute doesn't scale with length. Nisten speculated it could be a major step toward usable AI-generated movies by solving the face/character consistency problem.
ByteDance Seedream 3.0: bilingual 2K text-to-image model
ByteDance's Seed team announced Seedream 3.0, a powerful bilingual (Chinese/English) text-to-image model that generates native 2048x2048 images with fast inference of around 3 seconds for a 1K image on an A100. It challenges the top closed image generation models.
OpenAI launches GPT-4.1 family (4.1, mini, nano) in the API
OpenAI released the GPT-4.1 family of models, available via API only, in three sizes: 4.1, 4.1-mini and 4.1-nano. The family features a 1M token context window, in contrast to o3's 200k, and is aimed at developers building on long-context and coding workloads.
OpenAI open sources the MRCR long-context benchmark dataset
OpenAI open sourced MRCR, a benchmark dataset for evaluating long-context, complex retrieval tasks, building on Gemini research from Google and publishing the dataset on Hugging Face.
Meta drops Llama 4 Scout (109B) and Maverick (400B) open-weights MoE models
Meta released the long-awaited Llama 4 family in a chaotic Saturday drop: Scout (17B active / ~109B total, 16 experts) and Maverick (17B active / ~400B total, 128 experts), with a 2T-parameter Behemoth still in training. The models are multimodal, multilingual MoE architectures trained on ~30T tokens with FP8 and interleaved attention (iRoPE), claiming 10M context for Scout and 1M for Maverick. The release was marred by drama: the LMArena version differed from the released model, and the community criticized the lack of small local-friendly sizes.
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.
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.
Reve emerges with SOTA diffusion image generation claims
Reve launched a new diffusion image generation model claiming state-of-the-art quality, reportedly beating heavyweights like Midjourney and Flux at roughly a penny per image. The previously low-profile lab made a splash with strong prompt adherence and image quality.
AI21 releases Jamba 1.6 Large and Jamba 1.6 Mini open-weights models
AI21 Labs released Jamba 1.6 in Large and Mini sizes, updating its hybrid SSM-Transformer (Mamba-based) model family with open weights on Hugging Face. The Jamba architecture targets long-context efficiency compared to pure transformer models.
Inception Labs debuts Mercury, a commercial diffusion LLM
Inception Labs announced Mercury, billed as the first commercial-scale diffusion large language model, generating text via diffusion rather than autoregressive decoding. The approach promises dramatically faster token throughput, demoed first with the Mercury Coder playground.
Arc Institute and NVIDIA release Evo 2, a 40B state-of-the-art genomics model
Arc Institute and NVIDIA introduced Evo 2, a state-of-the-art genomics model with around 40 billion parameters trained on 9.3 trillion nucleotides. It uses the StripedHyena architecture to process genetic sequences up to 1 million nucleotides, enabling prediction of genetic mutation effects and even design of entire genomes. Fully open: two papers, weights, data, and training and inference codebases.
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.
Follow Model Architecture and everything else in AI — live every Thursday.
