open-source ai

Bilingual

Frontis-MA1: Open 35B model toward recursive self-improvement
a month ago
- OpenMLE is an open full-stack system for recursive self-improvement (RSI) in machine learning engineering (MLE), including environments (Gym), operator learning (RL), and long-horizon search (Evo).
- Frontis-MA1 (35B) is post-trained as a meta-evolution agent, using four atomic operators (Draft, Improve, Debug, Crossover) trained via execution-grounded SFT and RL.
- On MLE-Bench Lite, Frontis-MA1 improves Medal Average from 39.39% to 71.21% with OpenMLE-Evo-Max, surpassing GPT-5.5 + Codex and approaching GPT-5.6 Sol.
- Both the trained model and the search framework transfer to held-out NatureBench Lite, raising Match-SOTA from 50% to 70% and 20% to 50% respectively.
- The full stack—weights, gym, sandbox, training, search, and evaluation—is open-sourced for reproducible AI4AI research towards RSI.
I'm Sorry, Dave
2 months ago
- Claude AI refused to translate a blog post about Roma people and wolves, citing dehumanization, and offered to translate other content instead.
- The author compares this incident to HAL 9000's refusal in 2001: A Space Odyssey, arguing it represents ideological censorship.
- Contrast is drawn with Chinese AI model Kimi K2.7, which openly answered a question about Tiananmen Square, highlighting a double standard.
- The author advocates for strong open-weight models to prevent what they see as soft ideological tyranny from companies like Anthropic.
- The incident raises questions about AI safety and alignment: safety from what and alignment with whom when refusing harmless translations.
Kimi K3 Now Available via Telnyx Inference API
2 months ago
- Kimi K3 is a 2.8-trillion-parameter open-source model from Moonshot AI, available on Telnyx Inference API.
- It is the first open-source model in the 3-trillion-parameter class, featuring Kimi Delta Attention, Attention Residuals, a 1M-token context window, and native vision.
- The model competes with closed-source frontier models from Anthropic and OpenAI on coding, reasoning, and agentic benchmarks.
- K3 runs on Telnyx-owned GPU infrastructure and is accessible via an OpenAI-compatible API.
- Pricing: $0.27 per 1M cached input tokens, $2.70 per 1M input tokens, and $13.50 per 1M output tokens.
Kimi K3 after a week of real project use: what works, what doesn't
2 months ago
- Kimi K3 is the first open-source model capable of serious engineering conversations at the architecture level, focusing on system design and trade-off analysis.
- Its instruction following is on par with Opus, strictly adhering to boundaries and multi-layered instructions across long sessions.
- Data analysis quality matches that of Fable, providing meaningful observations and actionable insights from real datasets.
- For long coding sessions, K3 surpasses Opus and rivals Fable, but has post-training issues: overload errors, empty thinking tokens, and tool-calling logic loops that require robust error handling.
- Cache mechanism needs careful key management for cost efficiency, and subscription credit consumption is faster than Claude and Codex.
- Performance degradation with growing context is minimal, better than expected for an early-stage model.
- The broader context: discussions have shifted from whether Chinese open-source models are good to how to restrict them, indicating a significant technological threshold.
- K3's competitive pricing and frontier-level capability make it a serious emerging competitor to Claude and Codex, with the gap closing rapidly.
Mozilla: The state of open source AI
2 months ago
- Open-weight AI is being successfully adopted in diverse, real-world applications globally, from Māori language models to agricultural diagnostics, emphasizing ownership and control.
- A Mozilla/SlashData survey shows 79% of developers using AI choose open models, though only 51% of open-model teams reach production compared to 63% for closed, due to operational tooling and trust gaps.
- Open models have become a commercial market at a multi-hundred-billion-dollar scale, with companies like Databricks, Mistral, and DeepSeek achieving significant revenue and valuation growth.
- The strategic focus is shifting to owning layers above the model, such as the agentic harness (orchestration, tools, memory, permissions), where open versus closed competition is intensifying.
- Governments are increasingly supporting open AI, with policies like the European Commission's 'open source first' rule and Canada's target to boost business adoption from 12% to 60%.
- Key challenges include infrastructure costs, security concerns, maintenance, and deployment complexity, with regional variations highlighted in a global survey table.
- Risks to open AI's progress include a widening capability gap in reasoning and agentic tasks, potential dominance of closed platforms in permission standards, economic sustainability, and misuse concerns.
- The future direction advocates for competition, interoperability, and the freedom to switch vendors, warning against a 'rented' future controlled by a few, with a call to build openly and collectively.
German AI consortium releases Soofi S, an open 30B model
2 months ago
- A German research consortium released the open-source 30B language model Soofi S, trained entirely on Deutsche Telekom's AI cloud.
- Soofi S uses a resource-efficient Mixture-of-Experts (MoE) hybrid architecture, activating only 3.2B of its 31.6B parameters per token for constant processing speed with long inputs.
- The model outperforms other fully open models like Olmo 3 32B and Apertus 70B in German, English, and programming benchmarks, with a training mix deliberately weighted toward German data.
- Critics argued Soofi S was 'overtrained' by Chinchilla scaling law standards, but the technical leadership countered that those laws don't apply to MoE architectures due to data repetition benefits.
- Soofi S shows strong throughput in long-context generation, maintains near-constant performance from 4,000 to 256,000 tokens, and leads in German-specific knowledge tests while having some weaknesses in math and factual retrieval.
- The model was trained on sovereign infrastructure using renewable energy, with openness in releasing weights, code, and data inventory, though it doesn't fully meet a strict European open-data definition due to a small commercially licensed dataset.
Do frontier models matter if most production AI ends up running on open models?
2 months ago
- Chinese open-weight models accounted for 41% of Hugging Face downloads this spring, surpassing U.S. models.
- Open models, especially from Chinese firms, are handling a significant share of AI requests on platforms, suggesting a shift toward cheaper, customizable alternatives over frontier models.
- Hugging Face CEO Clem Delangue argues companies increasingly prefer owning AI models rather than renting them to avoid outsourcing core capabilities and to gain control.
- Fortune 500 firms are using Hugging Face to deploy private and open models, indicating a trend toward multiple specialized models rather than a single solution.
- Chinese AI labs regularly release capable open models, undercutting proprietary AI economics and competing with top models in areas like coding and security.
- Executives like Microsoft's Satya Nadella warn against single-provider lock-in, emphasizing the need for distributed learning infrastructure so firms control their data.
- Debate exists over open model risks, with concerns about misuse by bad actors versus arguments that concentration of power is a bigger danger.
- Delangue advocates for transparency in open models to level the playing field and allow defenders to address cybersecurity risks.

open-source ai

Frontis-MA1: Open 35B model toward recursive self-improvement

I'm Sorry, Dave

Kimi K3 Now Available via Telnyx Inference API

Kimi K3 after a week of real project use: what works, what doesn't

Mozilla: The state of open source AI

German AI consortium releases Soofi S, an open 30B model

Do frontier models matter if most production AI ends up running on open models?