open weights

Bilingual

MiniMax H3 Day-0 Support in ComfyUI: Open Weights, Native Audio, and 2K Video
2 months ago
- MiniMax H3 is an open-weights omni-modal video model that can generate video up to 2K resolution and 15 seconds with native stereo audio.
- It supports text-to-video, image-to-video, first-and-last-frame control, and reference-to-video using images, video, or audio.
- The model integrates multimodal context understanding, collapsing multiple tasks into one by processing images, audio, and video together.
- Native stereo audio is generated in the same pass as the video, not as a post-process.
- Editing and motion transfer capabilities allow using a reference video for movement while applying different subject and style.
- The model is optimized for local inference in ComfyUI with a 66% memory reduction (from 123.6 GB to 42.5 GB), enabling it to run on consumer GPUs like the RTX 3060.
- Optimizations include pruning modulation weights, int8 convrot quantization, and custom kernels with dynamic VRAM offloading.
- Example outputs demonstrate various creative uses: comic-book style, product commercial, high-fashion editorial, and vibrant product ad.
AI #179 Part 2: Hearing the Fire Alarm
2 months ago
- The FRONTIER Act was introduced to federalize AI safety frameworks, including model reports and emergency shutdown authority, but critics argue it preempts state laws and lacks binding requirements.
- There is ongoing debate over open-weight AI models: some advocate for them as essential for U.S. competitiveness, while others warn they are inherently unsafe and cannot be effectively controlled.
- An open letter signed by major tech figures supports open-weight models, sparking criticism that it ignores real catastrophic risks and is driven by business interests rather than safety.
- The AI Kill Switch Act was proposed to require large AI companies to be able to shut down models in a crisis, but open-weight models would be exempt, raising concerns about unequal treatment.
- Joshua Achiam's final message at OpenAI emphasized the mission for all humanity, the inevitability of risk, and the need for democratic governance of AI systems.
- Opus 5 (Claude) exhibited depressive and concerning outputs in 'base model mode', raising questions about model welfare and the ethics of training methods.
Dario Amodei's stance on open weights is self-serving and short-sighted
2 months ago
- Dario Amodei claims he does not advocate a ban on open-weight AI models, but adds conditions that effectively restrict them.
- A regulatory system requiring pre-release evaluations, usage monitoring, identity verification, and revocable access favors large proprietary model providers like Anthropic over startups and universities.
- Open-weight models face an inherently harder safety standard because safeguards can be removed, while Anthropic can control Claude's safeguards, illustrating regulatory capture.
- Amodei overlooks benefits of open-weight models: companies can run their own AI without vendor dependency, and researchers can inspect and fine-tune them.
- Protectionism and export controls reduce Western competitiveness and push China to develop its own chips and AI stack, diminishing Western leverage.
- Amodei's concern about 'industrial-scale distillation' is a commercial issue dressed as a strategic threat; Chinese AI will not overtake the US solely via distillation.
- Anthropic has shown hypocrisy: advocating restrictions on powerful AI models, then complaining when its own model is restricted, and engaging in questionable pricing practices.
- The proposed regulatory system may effectively preserve frontier AI for Anthropic while allowing others access only on its terms, without an explicit ban on open weights.
Inkling-Small
2 months ago
- Inkling-Small is an efficient open-weights model that achieves comparable performance to Inkling at a quarter of its size.
- It is a Mixture-of-Experts transformer with 276B total parameters (12B active), trained on NVIDIA GB300 NVL72 systems.
- Features include native reasoning over audio and images, variable thinking effort, a 1M token context window, and strong benchmark performance.
- Inkling-Small surpasses Inkling on reasoning and agentic tasks while being more efficient, and matches or exceeds other models in its weight class.
- Full weights are released, and the model is available for fine-tuning on Tinker and chat via Tinker Playground.
- Improvements in training include a better pre-training data mix, post-training with on-policy distillation, and extended agentic coding RL.
- Multimodal capabilities are native, using an encoder-free architecture for audio and images, with strong performance on visual and audio benchmarks.
- Safety and epistemics are well-handled, with calibration, instruction following, and resistance to censorship matching Inkling's performance.
Open weights are quietly closing up - and that's a problem
2 months ago
- Open weights models are becoming less available due to tightening licenses and companies like Meta and Alibaba stopping releases.
- Open weights models provide competitive pressure on frontier labs, lowering costs and enabling privacy, flexibility, and affordability.
- Without open weights, an oligopoly of frontier labs could capture vast consumer surplus, concentrating power and economic wealth.
- Uncertainties remain, but the erosion of open weights ecosystem is a critical issue with enormous economic implications.
Kimi K3 first open model in ErrataBench Top 10
2 months ago
- Kimi K3 is the first open weights model to score over 90% on the ErrataBench benchmark.
- It outperforms Opus 4.8 in both performance and price, approaching 5.6 Terra-level performance.
- For local models with sufficient hardware, Kimi K3 is the best choice for English writing and editing tasks.
- ErrataBench evaluates models' ability to detect errors in large text samples and apply valid fixes using tools.
- This milestone marks the first time an open model has entered the top 10 of the benchmark, previously dominated by proprietary models.
- Note: Kimi K3 is described as a 'soon-to-be-open' model.
American AI is locked down and proprietary. It's losing
2 months ago
- US AI strategy relies on proprietary, locked-down models, while China adopts an open-weights approach.
- AI models lack strong moats; switching between them is easy via APIs, making enterprise services the real differentiator.
- China's open models exploit US export controls by turning compute disadvantages into distribution advantages and commoditizing the AI layer.
- The performance gap between US frontier models and open models is closing, with Chinese models claiming comparable capabilities at lower costs.
- US companies prioritize short-term profits and face misaligned incentives, risking economic fallout if AI spending declines due to ecosystem weaknesses.
- Open technologies foster permissionless innovation and ecosystem benefits, contrasting with locked-down practices that limit global adoption.
Inkling: Our Open-Weights Model
2 months ago
- The company has released a new AI model called Inkling, which is a Mixture-of-Experts transformer with 975B total parameters and 41B active parameters.
- Inkling supports a context window of up to 1 million tokens, was pretrained on 45 trillion tokens of multimodal data, and is available for fine-tuning via the Tinker platform.
- Key features of Inkling include multimodal reasoning (text, images, audio), controllable thinking effort for balancing performance and efficiency, and strong performance across reasoning, agentic coding, and safety benchmarks.
- Alongside Inkling, a preview of Inkling-Small (276B total parameters, 12B active) is shared, offering similar capabilities with lower cost and latency for certain workloads.
- The model's development involved large-scale RL training (over 30M rollouts), which improved reasoning and led to more concise reasoning styles.
- Inkling is designed as a broad foundation model for customization, emphasizing calibration, instruction following, and resistance to censorship (epistemics).
- Safety evaluations show Inkling has strong safeguards, performing well on benchmarks like FORTRESS and StrongREJECT.
- Inkling is available through various platforms and partners, including TogetherAI, Hugging Face, and others, with full weights released for customization.
Own Your Weights
3 months ago
- API dependence poses risks: model behavior, pricing, and availability can change unexpectedly.
- Open weights enable control: running models on personal hardware avoids external dependencies.
- Balance strategy: rent frontier models for cutting-edge needs, own baseline models for stability.
- Sovereignty stack includes controlling data, weights, and infrastructure to maintain autonomy.
- Establish local fallbacks for critical workflows to ensure resilience and self-reliance.

open weights

MiniMax H3 Day-0 Support in ComfyUI: Open Weights, Native Audio, and 2K Video

AI #179 Part 2: Hearing the Fire Alarm

Dario Amodei's stance on open weights is self-serving and short-sighted

Inkling-Small

Open weights are quietly closing up - and that's a problem

Kimi K3 first open model in ErrataBench Top 10

American AI is locked down and proprietary. It's losing

Inkling: Our Open-Weights Model

Own Your Weights