Underrated AI Model Radar

Hy4 shows real OpenRouter traction and 20-50% lower costs despite its 770B MoE scale. Released August 28 with Apache 2.0 weights, it posts competitive...

Tencent Hy4's API delivers frontier coding performance at $0.834 input / $2.501 output per million tokens, with cheap cache hits at $0.042. No...

Tencent's 770B MoE model activates just 49B params per token with a native 10B MTP layer for speculative decoding. Open weights under Apache 2.0 and...

MiniMax M3 delivers free API access through Vercel and OpenRouter AI Gateway, giving developers an easy way to test its multimodal capabilities in coding, text, video and speech projects.

• Stealth test as Ox Alpha topped OpenRouter usage before reveal
• Free browser access now at chat.b.ai and chat.z.ai with no card needed
• MIT open...

CiteSure emerges as an underrated retrieval-augmented LLM that greatly outperforms GPT-4o and other baselines, with all datasets, models, and code freely available for local testing and deployment.

OpenRouter now acts as a unified gateway for large language models, adding support for embeddings, TTS/STT, and audio/video understanding across 500+ models. This broadens access to diverse capabilities in one place during launch week.

• Developer released dependency-free Go CLI that scores models on accuracy, instruction following, tool calling, and JSON output via OpenRouter.
  -...

Alibaba just opened Qwen3.8-Max globally for free testing on Token Plan, Qoder, and QoderWork ahead of next week's open-weights release of its 27B...

AgentSky delivers cloud-hosted agents supporting multiple harnesses (Claude Code, Codex, Hermes, OpenClaw) and LLMs, with one-click deployment and...

Kimi K3 delivers a 2.8-trillion-parameter open-weight model you can download and deploy locally at no cost, matching or beating proprietary giants on coding benchmarks while offering a 1-million-token context window.

agentOS delivers a free, open-source Linux OS as a lightweight library for AI agents, eliminating expensive sandboxes or VMs with 4.8ms cold starts...

Turbo Fieldfare streams Gemma 4 26B experts from SSD to run the full model in ~2 GB RAM on any M-series Mac, even 8 GB Airs, at 5.1–6.3 tokens/sec....

Memmy Agent is a free, open-source local memory hub that gives AI agents like Claude Code and Codex one shared, long-term memory across sessions and tools. MIT-licensed with 2M free tokens and four-layer architecture for contextual recall.

• Strong benchmark performance: The 350M LFM2.5-Encoder-350M scores 81.02 avg across 17 tasks, ranking 4th overall and beating several larger models
  -...

Moonshot AI just released its flagship Kimi K3 for free, ahead of open sourcing on July 27, creating a rare chance to experiment with a powerful model locally before wider availability.

Kimi K3 is now officially open weights with a good, fair license — a big win for open models. The 2.8T MoE model brings native visual understanding, a...

Andrew Ng's new OpenWorker stands out as a practical, underrated agentic framework that delivers finished work instead of just chatting.

• Handles...

• Free on Hugging Face after agreeing to Apache 2.0 license and Acceptable Use Policy, with contact info shared
• 8B model supports multimodal inputs...

New Open-Weight Releases

• 🔥 Cisco Antares: Cisco released Antares-350M and Antares-1B open-weight models on Hugging Face for repository-level...