Following opens with early access — sign up for early access to register your interest; it does not switch following on today. Everything you follow will land on one page, in your account menu.

1. Open Source & Self-Hostable✓ verified

   NVIDIA opened a proxy that lets one application speak both the OpenAI and Anthropic APIs

   Changing model or provider is normally an application rewrite; a translating proxy turns it into a routing rule you can measure both sides of.

   NVIDIA·2026-09-02
2. Open Source & Self-Hostable

   Hugging Face published 200-plus WebGPU kernels so models can run in the browser

   Inference in the browser is the cheapest deployment there is — no server and no per-token bill — and fast GPU operations across mismatched devices have been the missing floor under it.

   Hugging Face·2026-09-01
3. Open Source & Self-Hostable

   Tencent open-sourced Hy4 preview - 770B total parameters, 49B active, a 1M-token context

   A 49B-active open-weight model with a million-token window makes long-document work a self-hosting decision rather than a closed-API one.

   Tencent·2026-08-28
4. Open Source & Self-Hostable✓ verified

   Alibaba released Qwen3.8-Flash-Next, a sparse open-weight model that activates 6B parameters per token

   A 6B active-parameter model with a million-token ceiling is the shape that makes long-context agent work affordable on hardware you already own, and at a twelfth of flagship pricing it resets what a routine call should cost.

   Qwen (Alibaba)·2026-08-26
5. Open Source & Self-Hostable

   Z.ai shipped GLM-5.3-Flash, a 321B open-weight model built for cheap coding work

   Two large open-weight coding models landed on the same day at flash-tier prices; the cheap end of the coding-agent stack is now an open-weights question, not a frontier-lab one.

   Z.ai·2026-08-26
6. Open Source & Self-Hostable

   Ollama's 0.33 preview lets you switch local models on and off inside Claude's desktop app

   One misplaced system message was busting the cache on every request — worth checking how your own prompt is assembled before blaming the model for being slow.

   Ollama·2026-08-21
7. Open Source & Self-Hostable

   Ollama 0.33 makes a cancelled prefill resumable, and adds a Claude Desktop path

   If you self-host, a broken prefill restore is invisible until you look at latency; this is the release that stops a near-complete cache hit from re-running the whole prompt.

   Ollama·2026-08-21
8. Open Source & Self-Hostable

   llama.cpp starts a stable release channel with v0.2.0

   Packagers and anyone shipping llama.cpp inside a product finally have a version to pin that does not move with every commit.

   llama.cpp·2026-08-21
9. Open Source & Self-Hostable

   Ollama halves time-to-first-token by caching resolved model metadata

   Time-to-first-token is what a local model feels like, so halving it changes whether self-hosting is pleasant enough to use for real work.

   Ollama·2026-08-19
10. Open Source & Self-Hostable✓ verified

    Modular opens the Mojo compiler under Apache 2.0, a week after Mojo 1.0

    A closed compiler is a single point of vendor failure for anything built on it; Apache 2.0 on the compiler is what moves Mojo from an interesting runtime to something a team can commit a GPU codebase to.

    Modular·2026-08-18
11. Open Source & Self-Hostable✓ verified

    Qwen 3.8 27B scores 52 on the Artificial Analysis Intelligence Index, level with models far larger

    A 27B open-weights model at frontier-index parity is the size that actually fits on hardware you own — the point where running it yourself stops being a downgrade.

    Artificial Analysis·2026-08-17
12. Open Source & Self-Hostable

    FreeToken serves a 753B open-weight model from a single workstation GPU

    The gap between "open weights exist" and "I can run them" is a bandwidth-scheduling problem, and this is the clearest published attempt to close it on consumer hardware.

    arXiv·2026-08-17
13. Open Source & Self-Hostable

    llama.cpp folds --mmap, --no-mmap, --mlock and --direct-io into one --load-mode flag

    Anyone pinning llama.cpp in a Dockerfile, systemd unit or run script has a flag rename to make before the next bump — silent, and it fails at start-up rather than at build.

    llama.cpp·2026-08-15
14. Open Source & Self-Hostable✓ verified

    DeepSeek publishes V4 Pro 0813 weights at 1.7 trillion parameters

    A 1.7T open-weights checkpoint is not something most teams will self-host, but it sets the ceiling that hosted providers price against — and the missing license line is the thing to check before it reaches anyone's build.

    Hugging Face / DeepSeek·2026-08-13
15. Open Source & Self-Hostable

    NVIDIA ships a 30B Nemotron for agent sub-tasks plus an open router that cuts cost to roughly a third of Opus 4.8

    The stated tradeoff is published rather than hidden — 21–58% cheaper for 6–28% less accurate is a decision a team can actually take per sub-task instead of per product.

    NVIDIA·2026-08-11
16. Open Source & Self-Hostable✓ verified

    Meta releases Muse Glimmer, a 30B Apache-2.0 multimodal model built for local agentic use

    A 30B Apache-2.0 model that runs agentic work on your own hardware puts a real floor under what you have to pay a frontier API for — and llama.cpp already ships tool-call handling for it.

    Hugging Face·2026-08-10
17. Open Source & Self-Hostable

    MiniMax releases MiniMax-H3, an open-weights 33B image-to-video model — community tooling lands within a day

    Open-weights video generation at this scale moves a capability that was API-only last year onto self-hosted rigs — and the day-one tooling shows real demand.

    MiniMax·2026-08-06
18. Open Source & Self-Hostable✓ verified

    NVIDIA opens Alpamayo 2 Super, a 34B vision-language-action model for robotaxis, to commercial use

    An openly licensed driving-grade vision-language-action model lowers the entry cost for autonomy builders — and shows NVIDIA competing on models, not just chips.

    NVIDIA·2026-08-04
19. Open Source & Self-Hostable

    Mistral releases Shieldstral, a 3B open-weights policy-adaptive safety classifier

    Guardrails you can rewrite in plain language without retraining — and self-host on commodity hardware — change the cost calculus for anyone shipping user-facing AI.

    Mistral AI·2026-08-04
20. Open Source & Self-Hostable✓ verified

    Liquid AI releases LFM2.5-2.6B, an open-weights on-device model built for local agents

    Tool-calling agents that run entirely on a device change the privacy and cost math for shipping AI features — no per-token bill and no data leaving the machine.

    Liquid AI·2026-08-04
21. Open Source & Self-Hostable

    Cursor open-sources Mixture-of-Kittens, a deterministic MoE training megakernel for NVL72 racks

    An application company open-sourcing frontier-grade training kernels signals how far serious model infrastructure has spread beyond the big labs.

    Cursor·2026-08-04
22. Open Source & Self-Hostable✓ verified

    llama.cpp merges multi-token prediction for DeepSeek V4 — local decode roughly 50% faster

    Self-hosted inference of the newest hot open-weights family just got materially faster on consumer hardware — worth re-benchmarking your local stack.

    llama.cpp·2026-08-02
23. Open Source & Self-Hostable✓ verified

    LG ships K-EXAONE 2.0 — Korea's largest open model, 750B parameters under Apache 2.0

    Sovereign-AI programs are now shipping frontier-scale open weights under permissive licenses — the self-hostable ceiling keeps rising.

    LG AI Research·2026-07-31
24. Open Source & Self-Hostable✓ verified

    DeepSeek re-post-trains V4 Flash — and says the cheap model now beats its own flagship on agent work

    Post-training alone moved a 13B-active open model past a flagship on agent benchmarks — the cheapest tier is where agent workloads may now land.

    DeepSeek·2026-07-31