Skip to content

Multimodal

Beyond text: vision, mixed image-text, audio and video input and output in models and products.

Latest picks

81–100 of 592

Sep 24Thursday

Computing Life · Share · Yage

Qwen-Image-2.1: A version rollback that packs text rendering, editing, and native RGBA into one open-weight model

Qwen released Qwen-Image-2.1 on Sep 20, a 7B open-weight image model that unifies text-to-image, local editing, and native RGBA output in a single pipeline. The version number rolled back from 3.0 to 2.1 reflects a 2026 split: 3.0 is a closed-source commercial API, while 2.1 continues the open research branch. A built-in RGBA VAE outputs PNGs with transparency, skipping external matting. The interface supports up to 10 reference images and three mask types at native 2K. The license shifted from Apache 2.0 to a research-only agreement; commercial use requires a separate license. Community tests show ~25s per megapixel image on RTX 5070/5080 at 25 steps, ~15.6GB VRAM with Q8 quantization. Text rendering remains a strength, but multi-subject consistency shows facial generalization on well-known public figures—official demos don't guarantee universal performance.

Why it matters: The Qwen team open-sources Qwen-Image-2.1, folding text-to-image, local editing and transparent-background generation into one pipeline, with the version number dropping from 3.0 to 2.1.

The Verge · AI

Meta is bringing its Muse AI agent to smart glasses with voice activation

Two weeks after launching Muse, Meta says it's working on bringing the agent to its smart glasses, including the new ones shown at Connect. You'll activate it by saying its name and can ask it to guide workouts, log meals, or help shop for products you're looking at. The glasses are also getting an FDA-cleared hearing enhancement feature for adults with mild to moderate hearing loss. The post doesn't specify a launch date or which models will get Muse.

Why it matters: Meta says it is working on bringing its two-week-old Muse assistant to its smart glasses, where a spoken name triggers workout coaching, meal logging or price checks.

Sep 23Wednesday

AI HOT (Curated Pool)

Xiaomi releases open-source MiMo-V2.6 Pro and Flash multimodal models; Pro matches Claude Opus 5 and GPT-5.6 Sol on most agent benchmarks

Xiaomi open-sourced two multimodal models: MiMo-V2.6 Pro and Flash. Pro scored 46 on the Artificial Analysis Intelligence Index—the highest among open-source models—and matches Claude Opus 5 and GPT-5.6 Sol on most agent benchmarks. The post doesn't disclose parameter counts, training cost, inference latency, or the exact open-source license, so I'd hold off on production assumptions for now.

Why it matters: Xiaomi open-sourced the full-modality MiMo-V2.6 Pro and Flash, with Pro matching Claude Opus 5 and GPT-5.6 Sol on most agent benchmarks and scoring 46 on the Intelligence Index, the highest of any open model.

AI HOT (Curated Pool)

Qwen-Image-2.1 Tops Arena's Open-Source Leaderboards for Image Editing and Text-to-Image

Alibaba's Qwen-Image-2.1 ranks first among open-source models on Arena's Image Edit and Text-to-Image leaderboards. It scored 1367 in Image Edit Arena, placing 16th overall, just 3 points behind GPT-Image-1.5-high-fidelity at #15. The post doesn't disclose parameter count, architecture, or release timeline.

Why it matters: Qwen-Image-2.1 tops both the Arena image-editing and text-to-image open leaderboards, scoring 1367 in editing for 16th overall, three points behind 15th-place GPT-Image-1.5-high-fidelity.

Sep 22Tuesday

AI HOT (Curated Pool)

Qwen-Image-2.1 released as open weights, tops Image Edit Arena among open-source models

Qwen-Image-2.1 is out with open weights. It scored 1367 on the Arena Image Edit Arena, ranking #1 among open-source models and #16 overall — just 3 points behind GPT-Image-1.5-high-fidelity at #15. It also landed #1 open-source on the Text-to-Image Arena. The post doesn't disclose parameter count, architecture details, or the exact open license.

Why it matters: Qwen released Qwen-Image-2.1 weights, and the model sits first among open models on the Arena image editing board at 1367, three points behind GPT-Image-1.5 high-fidelity at 15th.

Latent Space

Xiaomi MiMo-V2.6-Pro tops open weights leaderboard, trained for $3M

Xiaomi released the MiMo-V2.6 series. The Pro version ranks #1 among open weights models on the Artificial Analysis Intelligence Index with a score of 46, at a training cost of $3M. A Flash variant targets efficiency, and an UltraSpeed variant offers 20x faster output. The technical report details RL scaling across three axes: larger batches and throughput, richer multi-task environments, and more grader compute. Code and training recipes are open-sourced, but the 7k+ task datasets are not yet released. Former DeepSeek engineer Fuli Luo, now at Xiaomi, previously live-streamed the training runs.

Why it matters: Xiaomi MiMo-V2.6-Pro tops the Artificial Analysis open weights intelligence index with 1.02T total and 42B active parameters, trained for about $3M, at $0.13 per task.

AI HOT (Curated Pool)

Xiaomi MiMo-V2.6-Pro hits ~10th on Code Arena WebDev, ~3rd among open-weight models

Xiaomi released two omni-modal models: MiMo-V2.6-Pro and Flash. The Pro version scored 1628 on Code Arena WebDev, up 153 points from MiMo-V2.5-Pro's 1475, landing around 10th overall and ~3rd among open-weight models under MIT license. The post doesn't disclose Flash's benchmark numbers or parameter counts.

Why it matters: Xiaomi's MiMo-V2.6-Pro scored 1628 on the Code Arena web development leaderboard, 153 points above the previous generation, around 10th overall and about 3rd among MIT-licensed open-weight models.

Sep 21Monday

AI HOT (Curated Pool)

Qwen-Image-2.1: A 7B Single-Checkpoint Model for Both Image Generation and Editing

Qwen-Image-2.1 is a 7B native image generation and editing model. It uses a single checkpoint for both tasks and supports up to 10 reference images. The model includes a built-in prompt-enhancement LLM, integrates with diffusers and ComfyUI, and offers a no-install browser demo on Hugging Face Spaces. The post doesn't disclose training data, inference latency, or benchmark comparisons.

Why it matters: Qwen releases Qwen-Image-2.1 at 7B, one checkpoint for both generation and editing, up to 10 reference images, already integrated with diffusers and ComfyUI.

Sep 20Sunday

AI HOT (Curated Pool)

Qwen-Image-2.1 now works with ComfyUI, open weights available

Alibaba Qwen released open weights for Qwen-Image-2.1 with native ComfyUI support. A single 7B checkpoint handles both image generation and editing, outputs up to 2K natively, accepts up to 10 reference images per instruction, and supports RGBA with alpha channel. The post doesn't spell out license terms or hardware requirements.

Why it matters: Alibaba open-sourced Qwen-Image-2.1 weights, a single 7B model handling both image generation and editing, with native ComfyUI nodes, 2K output and edits from up to 10 reference images.

Hacker News front page

StepFun launches Step 5 Preview, a 600B MoE flagship model targeting coding and finance

StepFun introduces Step 5 Preview, a 600B-parameter MoE model with 27B active per token, a 1M-token context window, and vision support. It scores 67.7 on DeepSWE v1.1, ahead of Kimi K3 and GLM-5.3 but behind GPT-6 Astra and Claude Opus 5. On the in-house StepCodeBench it hits 49.0, again leading domestic models and trailing the two US labs. On FrontierFinance it reaches 66.4, second only to Claude Opus 5. Artificial Analysis gives it an intelligence index of 44; StepFun claims substantially lower cost per task at comparable intelligence. The post does not disclose API pricing, release timeline, or training details.

Why it matters: StepFun released Step 5 Preview, a 600B-parameter MoE flagship aimed at coding and finance, scoring slightly above Kimi K3 and GLM-5.3 on public coding benchmarks such as DeepSWE v1.1.

Sep 18Friday

AI HOT (Curated Pool)

OpenRouter tested 20 image gen models: cheapest at $0.006, priciest at $0.134

OpenRouter sent the same prompt to 20 image models and read the actual billed cost. GPT Image 2 was cheapest at $0.006 per 1024×1024 PNG; Gemini 3 Pro Image was priciest at $0.134—a 22x spread. Pricing units differ across providers (tokens, megapixels, per image), so side-by-side list prices mislead; generate once and check usage.cost. Five of six models rendered text correctly, including the cheapest. Recraft V4.1 Vector outputs editable SVG at $0.08. The post also details formats, resolution caps, and seed support per model.

Why it matters: OpenRouter tested 20 image models with the same prompt and size, reading actual billing per response: cost per image runs from $0.006 to $0.134, a 22x spread, and billing units differ across vendors.

Sep 15Tuesday

Computing Life · Share · Yage

Runway demos code-free UI rendering; Google lets agents write their own manuals; GitHub Enterprise goes air-gapped

All three are early-stage. Runway Solaris generates interactive UIs frame-by-frame with no frontend code—only curated demos and a waitlist so far, no public testing, pricing, or API. Google WikiSkill distills agent failure logs into reusable skill manuals, lifting Gemini-3.5-Flash accuracy from 49.5% to 68.1%, but skills from a small model can hurt a larger one; no official code repo. GitHub GHES 3.22 lets enterprises self-host Copilot CLI inside air-gapped networks with admin-managed model endpoints, though many features are disabled and it's labeled a technical preview.

Why it matters: Runway's Solaris has a language model render interfaces frame by frame with a world model, and Runway reports 250 people preferred it to front ends from Claude Opus 5, with no third-party test yet.

Sep 14Monday

Hacker News front page

30 SVG prompts benchmark 2025–2026 LLMs on pelican-bicycle-style drawing tests

Tom Gally built a site with Claude Fable 5.1 that extends Simon Willison's pelican-riding-a-bicycle test into 30 SVG drawing prompts. The 2026 run covers six models—GPT-6 Astra, Claude Fable 5.1, Gemini 3.8 Flash, DeepSeek V4 Pro, Qwen3.8 Max, and Fugu Ultra v2—while the 2025 run includes ten models like Claude Sonnet 4.5 and GPT-5.1. Each image shows generation time and cost: DeepSeek V4 Pro finished in 1 min 36 s at $0.10, Qwen3.8 Max took over 12 minutes, and Fugu Ultra v2 cost $1.02. The post presents raw SVG outputs without subjective ratings, so you compare the drawings directly.

Why it matters: Tom Gally ran 30 SVG drawing prompts against 2025 and 2026 model batches and shows the raw outputs side by side, so the same prompt can be compared across models.

Sep 12Saturday

Latent Space

DeepSeek V4.1-Flash: a 763B encoder-decoder MoE with 8B prefill, 16B decode, and native vision

DeepSeek dropped V4.1-Flash on Sep 10. Despite the 4.1 label, Sebastian Raschka called it a V5-level rewrite. It's a 763B total-parameter MoE with a causal encoder-decoder split: 8B active for prefill, 16B for decode, yielding 1–2% sparsity and up to 8× smaller KV cache vs V4 Flash. Native vision is built in, and V4 Pro has been quietly retired. The post doesn't include benchmark tables but argues current evals miss the point—the real advance is context efficiency for long-running agents.

Why it matters: DeepSeek's V4.1-Flash splits prefill and decode into 8B and 16B activation sets with a causal encoder-decoder and native image input, making it primary material for reading the architecture tradeoffs.

AI HOT (Curated Pool)

DeepSeek V4.1-Flash open-sourced: CED architecture cuts prefill cost for coding agents

DeepSeek released open weights for V4.1-Flash, a 552B MoE model with a Causal Encoder-Decoder architecture tuned for coding agents. It splits compute asymmetrically: 8B active params during prefill, 16B during decode, plus improved KV cache efficiency. On Terminal Bench 2.1 it hits 90.6; on Automation-Bench it scores 54.8—better than V4-Pro but still failing roughly half of complex workflows, so keep a human in the loop. It is also DeepSeek's first non-experimental model with native image input. Chartography reaches 78.9, but ZeroBench logical reasoning over images is only 49. DeepSeek has already retired V4-Flash traffic and will reroute V4-Pro traffic to V4.1-Flash starting September 14.

Why it matters: DeepSeek open-sourced V4.1-Flash, a 552B multimodal MoE model with a Causal Encoder-Decoder architecture that activates 8B parameters at prefill and 16B at decode, aimed at cutting prefill cost for coding agents.

Sep 10Thursday

AI HOT (Curated Pool)

DeepSeek-V4.1-Flash lands on SiliconFlow, a 552B MoE with 1M context window

SiliconFlow launched DeepSeek-V4.1-Flash on Day 0. It's a 552B MoE model with ~8B active params during prefill and ~16B during decode, native vision, and a 1M-token context window. KV cache footprint is about 1/4 of V4 Flash, which helps with deployment cost. MIT license keeps commercial use straightforward.

Why it matters: SiliconFlow launched DeepSeek-V4.1-Flash the same day, a 552B-parameter model with 1M context that cuts KV cache use to a quarter of V4 Flash and lowers deployment cost.

AI HOT (Curated Pool)

DeepSeek V4.1-Flash cuts KV cache memory for AI agents to a quarter of its predecessor

DeepSeek released V4.1-Flash, a 552B-parameter model built to slash memory costs for AI agents. Its KV cache in fast GPU memory is about a quarter the size of V4-Flash, and the offloaded portion shrinks to roughly an eighth. The model splits into an encoder and decoder: only 8B parameters activate per token during input processing, versus 16B during text generation, nearly halving input compute. It supports 1M-token contexts and stores the main KV cache in FP4. On the DeepSWE v1.1 coding benchmark it scores 74.2%, narrowly beating Anthropic Opus 5 and OpenAI GPT-5.6 Sol, but it still trails on complex scientific tasks and image analysis. Weights are on Hugging Face under the MIT license. The post does not disclose inference latency or specific hardware requirements.

Why it matters: DeepSeek released V4.1-Flash, cutting the KV cache that long-context agents need to about a quarter of the previous generation and halving data-input compute, with concrete architecture numbers for anyone weighing long-context deployment cost.

AI HOT (Curated Pool)

DeepSeek V4.1-Flash drops with native vision and a big price cut

DeepSeek released V4.1-Flash with a new Causal Encoder-Decoder architecture and native vision understanding — no separate vision model needed. It's a 552B MoE, activating 8B params for input and 16B for output. The post doesn't disclose the exact price cut or benchmark numbers, so I'd wait for third-party evals before getting excited.

Why it matters: DeepSeek released V4.1-Flash, which builds visual understanding into the model through a causal encoder-decoder architecture, activating 8B input and 16B output parameters out of 552B total.

AI HOT (Curated Pool)

DeepSeek Releases V4.1-Flash: New Causal Encoder-Decoder Architecture with Native Vision

DeepSeek V4.1-Flash is the smallest model in the new architecture family: a 552B MoE with 8B active params for input and 16B for output. It uses a Causal Encoder-Decoder design with native vision. KV cache drops to 1/4 of HBM and 1/8 of SSD storage vs the previous generation, and API pricing is lower. The post doesn't disclose exact pricing or vision benchmarks.

Why it matters: DeepSeek's smallest V4.1-Flash adds native vision and cuts KV cache needs to a quarter of the last generation's HBM, at a lower API price.

DeepSeek · API updates

DeepSeek-V4.1-Flash ships with native vision and a lower API price

DeepSeek released DeepSeek-V4.1-Flash, the smallest model in its new architecture series, with native multimodal vision understanding. It is live on the DeepSeek API under deepseek-flash, and API prices dropped at the same time. The older V4 Flash and V4 Flash Vision Exp are retired, with their call names routed to V4.1 Flash for now.

Why it matters: DeepSeek routes the old model names to V4.1 Flash and cuts API prices, so existing calls keep their names but switch models.