ten models, vision included — and why ours run at 2×
The catalogue just grew from seven models to ten, gained image input, and — on one model — doubled in speed without changing a single line of engine code. The new arrivals are worth introducing. So is the discipline behind that 2×, because it's the actual product.
the three newcomers, and who they're for
GLM-4.7-Flash — the one your legal team will approve. A 31B mixture-of-experts model from Z.ai under MIT — the most permissive licence in our catalogue, with no usage policy riders to send to counsel. It's also, plainly, quick: 118 tokens/second measured on a single consumer GPU, the fastest decode in the pool. Strong general reasoning, 200K context, from €0.16 per million input tokens.
Qwen3 Coder 30B — code as a first-class workload. The most-downloaded open coding model of its size (12M+ downloads), built for repository-scale work: 262K context, tool use, and a mixture-of-experts design with ~3B active parameters, so it answers at conversational speed rather than making you watch a large model think. Measured at 96 tok/s. Apache-2.0.
Qwen3.6 35B-A3B — the delegate. If you build agent systems, you know the shape: a strong model plans, and a fast model executes the sub-tasks. This is the fast one that's still smart — the full 35B answers, but only ~3B parameters activate per token. 76 tok/s measured, 262K context, and it accepts images. Apache-2.0.
vision, priced like text
Qwen3.8 27B — our flagship — and Qwen3.6 35B now take images through the standard OpenAI multimodal format: your SDK already speaks it. Two design choices worth knowing:
A text-only model refuses an image with a clear 400 instead of silently dropping it and confidently answering a question it never saw — an answer you'd have paid for and could never have detected. And images are inlined (data: URLs) rather than fetched from links, because your job runs on a community machine and following a URL would make that machine fetch arbitrary addresses on your behalf. What you send is exactly what the model sees. No separate image fee — normal token pricing.
The projector weights that give a model sight are pinned to the same upstream commit as the model itself and verified by SHA-256 like everything else on a host. A model that can see says so in GET /v1/models; the claim and the weights are checked against each other in CI, so neither can drift.
the 2×: same weights, same engine, one better decision
Here's the part that matters beyond this week's catalogue. Modern open models often ship with a multi-token prediction head — a small extra network that lets the model draft several tokens ahead and verify them in one pass. When the weights carry it, decoding gets dramatically faster. When they don't, no flag can conjure it.
And here's the trap: the same publisher frequently ships the same model in two repositories whose names differ by three letters — one where that head survived the conversion to the serving format, one where it didn't. The model card won't tell you. The config file will even claim the head exists when the shipped weights dropped it.
We measured it directly on Qwen3.6 35B, same quantization, same engine, same GPU: 33 tok/s without the head, 76 with it. So we built the check into our pipeline: before any build is pinned into the catalogue, we read the actual tensor list out of the published file — a two-second ranged request, no 16 GB download — and confirm what the weights really contain. Every speed we advertise is measured on the exact build we serve, on consumer hardware, and printed in the changelog with the card it came from.
That's the catalogue's real promise. Anyone can list open models. Choosing the build of each one that actually earns its VRAM — and proving it with numbers — is the work, and it's why a request to qwen3.6-35b-a3b means the same verified weights on every machine in the pool.
one model id, any card
Each catalogue entry spans up to nine quantizations, from ~11 GB cards to 24 GB flagships. Your request names the model; the network picks the best build the serving card holds, and the receipt records — cryptographically — which one ran. Quality tiers are priced apart, so a compact build costs less than the reference one, and you can pin a minimum tier per request when it matters.
The full list, with live per-token prices, context and capabilities: elephantpool.ai/models — generated from the same data the API serves, so the page can't drift from the product.