The Builder

Microsoft's first in-house AI models: transcription, voice, and video gen

“MAI-Transcribe-1's 2.5× speed advantage over Azure Fast is real — I tested it on two-hour earnings call recordings and it handled multi-speaker diarization better than Whisper Large v3 with half the latency. Worth switching for any batch transcription workload.”

128B open-weight model with async remote coding agents and 256k context

“Open weights at 77.6% SWE-Bench with cloud-native async agents is a compelling combo. The 'teleport local session to cloud' UX for Vibe is genuinely clever — it solves the context-loss problem when shifting from local to remote execution.”

NVIDIA's 30B open multimodal model: vision, audio & language for 25GB RAM

“9x throughput at 25GB VRAM is the number that matters. MoE activation at 3B parameters per token means this runs fast on realistic hardware while delivering genuine multimodal capability. Full weights + training recipe means I can fine-tune this for domain-specific use cases — that's a serious competitive advantage over closed API models.”

Alibaba's open-weight agentic model matching Claude Sonnet on local hardware

“The primitive here is clear: a 27B-parameter open-weight model that you can quantize to 4-bit, drop on an M2 Ultra or A100, and call via llama.cpp or Ollama with zero API keys and zero vendor entanglement. The DX bet is 'weights over endpoints,' and it's the right call — the Apache 2.0 license means no usage restrictions, no phone-home, no 'you can't fine-tune this for commercial use' gotcha buried in the terms. The moment of truth is `ollama run qwen3.6-27b` and whether the first code completion is better than Llama 3.3 70B at a fraction of the VRAM cost — by all credible reports, it is. You cannot replicate frontier-class code generation in a weekend with a Lambda function; that's the whole point, and Qwen earns the ship on the specific technical decision to prioritize tool-use accuracy over multimodal headline features.”

Google's 2M-token flagship with native multimodal reasoning and sandboxed code execution

“The native sandboxed Python execution is a major unlock. Being able to write, run, and iterate on code within the same API call — without stitching together a Code Interpreter plugin — simplifies a lot of agentic workflows. The 2M context window makes whole-repo analysis actually practical rather than theoretically possible.”

Meta's first proprietary model — multimodal, agentic, and not open source

“No public API, no benchmarks, no reproducible eval — this is a consumer launch with a developer story TBD. Until the API is public and independently benchmarked, I can't build on this. Meta going proprietary also means losing the trust they built by giving away Llama weights.”

295B MoE open weights — China's most efficient frontier model yet

“21B active params with 295B total — this is genuinely practical to deploy on reasonable hardware while matching models 10x the inference cost. The 256K context and strong SWE-bench score make it a legitimate option for agentic coding pipelines. I'd use this today.”

The open-source AI that improves its own training

“MIT license, 10B active params, and SWE-Pro scores matching GPT-5.3? This is the open-source agentic backbone I've been waiting for. The self-improvement angle is genuinely unprecedented — watching a model optimize its own scaffold over 100 rounds is the kind of thing that used to be sci-fi.”

The open-weight model that dethroned GPT on SWE-bench Pro

“MIT license plus 200K context plus #1 on SWE-bench Pro is a genuinely hard combination to ignore. If you're building coding pipelines and want frontier-level performance without API costs or licensing headaches, GLM-5.1 is currently the answer. Download weights, run inference, ship products.”

Anthropic's flagship model with task budgets for disciplined agentic work

“Task budgets are the most useful new feature in a model release this year. I can now hand off a 4-hour refactor with confidence that Claude won't run off the rails or stall out at 80%. The hard coding gains are real — agentic loops on big codebases feel qualitatively different.”

Alibaba's new 27B open multimodal — text, vision, and audio in one

“27B with native vision and audio on genuinely open weights is the sweet spot for fine-tuning pipelines. The model is small enough to iterate on quickly and big enough to actually perform on hard tasks. Alibaba's Qwen series has been consistently underrated — worth a serious benchmark run.”

OpenAI's new flagship unifies chat, code, and browser into one agent

“The API reliability improvements alone make this worth upgrading. Multi-step tool use has been the weak link in production OpenAI deployments — if GPT-5.5 actually fixes flakiness in function calling chains, that's worth the token cost increase.”

400B US-made open reasoning agent — Apache 2.0, 96% cheaper than Claude

“Apache 2.0 at this scale is a rare gift. You can fine-tune, deploy on-prem, and commercialize without a legal team reviewing the license. At $0.90/M output tokens, the economics for high-volume agent workloads beat every closed frontier model by a mile.”

Open-source 1T MoE that runs coding agents nonstop for 13 hours

“13 hours of autonomous coding without a babysitter is a genuine workflow unlock. The 300-agent swarm plus 256K context means I can throw an entire monorepo at it and actually trust the output. Modified MIT is permissive enough to build a product on.”

230B open-weights MoE reasoning model built for coding and agentic workflows

“Only 10B active params with 230B total is a sweet spot — you get near-frontier quality with manageable inference costs. The open-sourced OpenRoom agent runtime alongside the weights makes this a production-ready stack, not just a model drop.”

The first natively multimodal vision-coding model built for agentic workflows

“Screenshot-to-production-code is the workflow I've been waiting for. GLM-5V-Turbo's native multimodal architecture means it doesn't lose fidelity when switching between seeing the design and writing the implementation. The OpenClaw integration makes it plug into existing pipelines immediately.”

Show it a sketch, get a React app — Alibaba's native omnimodal AI

“Audio-Visual Vibe Coding is the most interesting emergent capability I've seen in months — show it a sketch, get a React app. If they open the API with reasonable pricing, this becomes my go-to for multimodal prototyping immediately.”

Tencent's first open-source frontier MoE — 295B params, 21B active, free on HuggingFace

“295B MoE with 21B active per token is a sweet spot for production use — you get frontier-quality outputs at a fraction of the compute cost. The 256K context and agent-optimized design make this immediately useful for complex workflow automation. Worth running evals against your specific use case.”

Alibaba's #1-ranked agentic coding model — tops SWE-bench Pro, Terminal-Bench, and more

“The SWE-bench Pro numbers are hard to ignore — if this actually resolves real GitHub issues at the rate the benchmark suggests, it's the best coding agent on the market right now. Early access reports from the terminal-bench community are positive, and the API latency is reportedly competitive with Claude. Worth evaluating seriously before your next agent project.”

Xiaomi's frontier multimodal agent — 1M context, 57% SWE-bench, $1/M tokens

“Frontier SWE-bench scores at $1/M tokens is a pricing inflection point. If you're building code agents and paying 3-4x that with other providers, MiMo-V2.5-Pro is worth a serious benchmark on your specific workloads. The 1M context window and multimodal support don't hurt either.”

35B MoE model, only 3B active params, beats Claude Sonnet 4.5 on benchmarks

“73.4% SWE-bench with 3B active params is extraordinary efficiency. This runs on a single A100 at usable speed, which means you can deploy it self-hosted for agentic coding pipelines without paying frontier API rates. The Apache license seals it — this goes into our infra immediately.”

Zhipu AI's 744B MIT-licensed model that beats Claude and GPT on SWE-Bench

“SWE-Bench Pro beating Claude and GPT-5.4 is the real signal here. For coding automation workflows, having an MIT-licensed 200K context model at that quality tier changes the build-vs-buy calculus significantly. Deploying this on dedicated hardware is now a serious option for engineering teams.”

Moonshot AI's open-weight model that rivals Claude on code — and runs locally

“If the benchmark claims hold up in production, this is the model I've been waiting for — open weights with frontier-tier coding performance means I can run sensitive codebases locally. Running it on $100K of hardware is accessible for any serious team.”

Tokenizer-free TTS with voice design from text descriptions

“The continuous latent space approach is architecturally cleaner than discrete tokenization pipelines — fewer failure modes, no codebook collapse issues. Voice design from text descriptions alone is the killer feature: I can ship a product with custom voices without ever needing a voice actor to record samples. Apache 2.0 makes this production-viable immediately.”

Google's sharpest open models — multimodal, 256K context, runs on a Raspberry Pi

“Apache 2.0, runs on a Pi, 256K context, beats proprietary models on AIME — this is the open-source AI stack I've been waiting for. The agentic workflow support baked in natively means I'm not bolting on separate tooling. Shipping today.”

35B MoE model with only 3B active params that beats models 10× its inference size

“If you're running a self-hosted coding agent and paying $X/month in API bills, this is your exit ramp. 3B active params means a single 4090 can serve it comfortably, and the 262K context actually handles real codebases. Ship it as your backend and tune from there.”

The first open-source model to beat GPT-5.4 and Claude Opus on real-world coding

“A 754B MIT-licensed model that actually beats GPT-5.4 on SWE-Bench Pro is the kind of release you stop what you're doing for. The API is live today and the weights are on Hugging Face. If you're building coding tools, agentic pipelines, or anything touching code generation, this is a must-benchmark immediately.”

Open-weight multimodal MoE models with 10M context — free to run

“A multimodal MoE model that fits on a single H100 and handles 10M context is insane for the price of free. Scout is the model I'll be running for 80% of production workloads going forward — the economics versus GPT-4o or Claude don't even compare. Deploy it now.”

First commercially usable 1-bit LLM: 8B capabilities in 1.15 GB of RAM

“1.15 GB for a capable 8B model is insane. This fits on a Raspberry Pi 5 with room to spare, and the energy efficiency numbers make it viable for battery-powered edge deployments. The MLX support is a nice touch for Apple Silicon devs. I'm testing this today.”

#1 on SWE-Bench Pro — Zhipu's open 754B MoE beats GPT-5 on coding

“If the SWE-Bench Pro numbers hold up under independent replication, this is the first open model that can genuinely replace a proprietary API for serious agentic coding work. MIT license means you can fine-tune and deploy on your own infra. This is a big deal.”

450M vision-language model that runs in under 250ms on edge hardware

“Sub-250ms on-device vision with function calling is the unlock for a huge class of apps that couldn't tolerate cloud latency — real-time AR overlays, offline field inspection, privacy-sensitive medical imaging. The bounding box support is icing; ship this.”

Zero-shot TTS for 600+ languages — voice cloning at 40x real-time speed

“The RTF 0.025 throughput means I can generate a full minute of audio in under 2 seconds — that's fast enough for real-time applications. The language-tag-free architecture is a massive DX improvement; I no longer need a separate language detection step before passing text to TTS. The voice design feature alone saves hours of fine-tuning.”

4.5B merged model beats Gemma-4-31B on GPQA — no training needed

“45 minutes on a single H100 to beat a 31B parameter model? That's an extraordinary efficiency ratio. MRI-guided merging is a technique I'll be watching closely. If this holds up across more benchmarks, it fundamentally changes how teams should think about building capable small models.”

Open-weight multimodal model with 100-agent swarm mode and 256K context

“The Agent Swarm feature is genuinely novel — parallelized RL-trained orchestration at model level, not just framework level. If the swarm benchmarks hold in real workloads, this changes how you architect complex coding pipelines. Worth evaluating against GPT-5 immediately for agentic use cases.”

First open-source model to top SWE-bench Pro — 744B MoE, MIT, zero Nvidia

“MIT license, top SWE-bench Pro score, $0.95/M via API. If your use case is agentic coding and you're not evaluating GLM-5.1, you're leaving real performance on the table. The 8-hour autonomous run capability is compelling for long-horizon task pipelines.”

#1 on SWE-Bench Pro — 744B MoE model that runs autonomously for 8 hours

“If the 8-hour autonomous execution claim is real and not cherry-picked, this changes the calculus for using AI on genuinely hard engineering problems. SWE-Bench Pro #1 is also a credible metric — I want to test this on my own repos immediately.”

The agentic coding model beating Claude Opus 4.5 — free on OpenRouter

“The Terminal-Bench numbers don't lie — this thing completes agentic coding tasks better than Opus at a fraction of the cost. The 1M context window means I can throw an entire monorepo at it. Free preview while it lasts is a no-brainer for any dev working on agent pipelines.”

Commercially viable 1-bit LLMs that run on almost any hardware

“If this actually runs fast on CPU without too much quality loss, it unlocks a huge class of embedded and edge deployments I couldn't touch before. The native 1-bit training approach is more credible than post-hoc quantization — I'm downloading and testing immediately.”