Best offer on record: SGLang open-source inference engine (Apache-2.0) — permanently free. 2 verified findings, checked 2026-10-22.
Offers
SGLang open-source inference engine (Apache-2.0) — permanently free high confidence
SGLang is a fully open-source LLM/VLM/diffusion serving engine (RadixAttention prefix caching, structured generation, speculative decoding) under Apache-2.0: $0 license, no subscription, no per-token/seat fees, unlimited usage, indefinite. 'Free' here means the software; you still pay for your own GPU compute (e.g., ~$3.39/hr median on-demand H100 as of Sep 2026). The commercial backer RadixArk (launched May 2026, $100M seed at $400M valuation) sells managed services, but the open engine remains free. No free tier, trial, education license, or startup program exists on SGLang's official site (verified sglang.io 2026-10-22: no pricing page, no student/startup program; only 'enterprise adoption/consulting/sponsorship' contact at sglang@lmsys.org).
Eligibility
Anyone, anywhere — no .edu email, no startup age/funding cap, no credit card, no region limits. Open-source license is unconditional (Apache-2.0). Students get the identical free software as enterprises.
How to claim
git clone or pip install/uv sglang (or Docker image) from https://github.com/sgl-project/sglang; run on any supported GPU/TPU/NPU/CPU (A100/H100/H200/B200/MI300X/TPU v6e-v7/Apple Silicon etc.). No account, no signup, no key.
Expiry
Permanent (open-source license; no expiry). Ecosystem sub-projects (SGLang Diffusion, SGLang-Omni, SpecForge, HiCache/Mooncake) are in the same repo/organization and open-sourced too.
Gotchas
The $0 software does NOT cover compute: GPU rental is the real cost (on-demand H100 ~$3.39/hr; below ~2.5-3B tokens/mo a per-token API is usually cheaper). Model weights for served models come from other labs — some models (e.g., certain gated LLMs) require their own license acceptance. Third-party 'pricing' sites (costbench, aiindigo, aitoolsatlas) describe a '$0 free tier' but aitoolsatlas's 'education pricing available' claim is generic boilerplate, not a real SGLang program. The 'median contract $51,500/yr' figure on costbench refers to RadixArk commercial managed deals, not the free engine.
AMD Developer Cloud free MI300X credits (~$100 / ~50 hrs) via SGLang partner guide medium confidence
AMD Developer Cloud has offered free MI300X GPU access to developers, and SGLang's own project posted (2026-03-17) a guide to run a full stack (Qwen3.5-122B-A10B-FP8 served via SGLang in one Docker command, selectable in OpenClaw's onboarding CLI) 'at zero cost' on a single MI300X, quantified as ~$100 AMD Developer Cloud credits (~50 hrs of MI300X). AMD replied to the post confirming the guide. This is partner-sourced free compute, not an SGLang program.
Eligibility
AMD Developer Cloud account holders (developer sign-up); guide is aimed at any developer/student/startup building an agent stack — no .edu requirement stated, no funding cap. US-based cloud (AMD data centers); account verification typically required.
How to claim
Sign up at devcloud.amd.com, apply for the free MI300X trial/credits, then follow SGLang's published guide (shared 2026-03-17 on the SGLang LinkedIn, shortened link lnkd.in/g2UZFehh) to deploy Qwen3.5-122B-A10B-FP8 via SGLang in one Docker command.
Expiry
Credits are promotional and time-boxed; ~50 hrs / $100 figure is from the Mar 2026 post — exact expiry window not confirmed (AMD's pages are JS-only and I could not read official terms today; search budget exhausted).
Gotchas
Not an SGLang-owned program — terms/limits/credits can change or be revoked by AMD; capacity on MI300X instances is limited and often queued; data egress/model-download costs may apply; 'free' applies only while credits last, after which MI300X time is metered. Verify current credit amounts at devcloud.amd.com before relying on this.
No SGLang-specific free offer was sunset or materially changed in 2025-2026: there is no SGLang free tier, trial credit, startup program, or education license to sunset — the 'free offer' is the Apache-2.0 software itself, which is permanent. Material 2025-2026 context: (1) the commercial entity RadixArk launched May 2026 ($100M seed, $400M valuation) with managed-inference/managed-training services on top of the open core — at this date it has announced no startup or education program, so nothing free exists on that front to claim; (2) SGLang joined the PyTorch ecosystem during 2025; (3) a free-compute path via partner AMD Developer Cloud (MI300X, ~$100/~50 hrs) was highlighted by the SGLang project in March 2026, but AMD's official terms could not be read (JS-only pages) and web-search budget was exhausted, so treat that figure as unverified current. Third-party discount sites claiming SGLang 'education pricing' (e.g., aitoolsatlas.ai) are auto-generated generic text, not a real program.