Two thirds less memory, wherever you run it.
The base is Qwen3.6-35B-A3B, compressed from 66.9 GB to 24.5 GB. It is a mixture-of-experts, so about 3B parameters are active on any one token and it answers at the pace of a far smaller model.
Every model we compress, measured the same way.
Each one is measured against the same base model compressed a different way, on the same corpus. Only the compression differs, so the comparison is like for like.
Tileward 35B-A3B: Qwen3.6-35B-A3B, compressed to a third
The model serving requests today. A mixture-of-experts: 256 experts per layer, 8 of them consulted on any one token. It costs about what a 3B model costs to run, and knows what a 35B model knows., so it answers at the cost of a much smaller model while holding the knowledge of a much larger one.
| Measure | Result |
|---|---|
| Base | Qwen3.6-35B-A3B |
| Parameters | 35B total, ~3B active |
| Size on disk | Weights on disk, measured. The expert layers, which are nearly all of the weight, go to 4 bits on a group of 32 rather than the usual 128. Attention, the router, the embeddings and the output layer stay at full precision, which is why the ratio is 2.74× and not higher. |
| Smaller by | 63%, a 2.74× ratio |
| Context window | The model is built for 262,144 tokens. The 65,536 figure is the limit our hosted deployment runs at, a serving choice about memory per request, not a property of the weights. Self-hosted you set your own. |
| Per-tile locking | Available |
A drop-in for FP8: the same score, 1.46× smaller
Both arms are the same base model on EleutherAI's lm-evaluation-harness, full MMLU: 14,042 questions, five-shot, log-likelihood over the four answer choices, no sampling, so the run repeats exactly. One machine, one set of settings for every arm., so the only thing that varies is the compression. Closeness is the point: a build that scores differently, in either direction, is behaving differently in ways nobody can account for afterwards.
| Build of Qwen3.6-35B-A3B | MMLU 5-shot, higher is better |
|---|---|
| Publisher's official FP8 | 82.60% ±0.31 |
| Tileward 35B-A3B | By category: humanities 77.77, social sciences 89.76, STEM 80.40, other 84.97. Two community 4-bit builds of this base score above us, GPTQ at 83.30% and AWQ at 83.17%. The full four-way table is on the evals page. |
The trade is against FP8, and sameness is the product. Two hundredths of a point apart on an error bar of three tenths is A 0.02-point difference against a ±0.31 margin carries no signal. MMLU is one multiple-choice benchmark, so it is consistent with the two builds behaving alike rather than proof of it. The direct test, per-token agreement against the FP8 build’s own outputs, is not something we have run yet., which is exactly what a substitute should look like. Same base, same tokenizer, same answers as far as we can measure, at 24.5 GB of weights against its 35.7 GB. If you serve FP8 today, nothing above the model has to change.
Two community 4-bit builds score higher, and that is not the same as better. GPTQ reaches 83.30% and AWQ 83.17%, both above us. But a build whose answers move on this benchmark has moved on the ones nobody ran, in directions this benchmark cannot tell you. If your prompts and evals are already tuned against the base model, a different score is a change to absorb, not a gift. If they are not, those builds are worth your time and we would rather you knew. The full four-way run.
What it costs to call. Metered per token on the hosted API, at this model’s own published rate. Self-hosted it is the licence and your own hardware, with no metering. See pricing.
gpt-oss 120B: served exactly as published
Not every model on the API is one we compressed. This one is third-party open weights under Apache 2.0, served as its publisher shipped it, and we have changed nothing about it.
| Measure | Result |
|---|---|
| Parameters | 117B total, ~5.1B active |
| Weights | As published, unchanged |
| Compressed by Tileward | No |
| Context window | The model is built for 262,144 tokens. The 65,536 figure is the limit our hosted deployment runs at, a serving choice about memory per request, not a property of the weights. Self-hosted you set your own. |
| Per-tile locking | Not available on this model |
| Where it fits | High-volume work where a smaller active path is enough |
There is no quality or size claim of ours on this panel, because we have not earned one. We did not compress this model, so any number here would be the publisher's, not ours. It is on the API because it is useful and cheap to run at volume, and it is in this list because you can call it, not because we changed it. The compression comparisons on this page are about Tileward 35B-A3B.
gpt-oss 20B: served exactly as published
Not every model on the API is one we compressed. This one is third-party open weights under Apache 2.0, served as its publisher shipped it, and we have changed nothing about it.
| Measure | Result |
|---|---|
| Parameters | 20B total, ~3.6B active |
| Weights | As published, unchanged |
| Compressed by Tileward | No |
| Context window | 8,192 tokens |
| Per-tile locking | Not available on this model |
| Where it fits | High-throughput document ingest |
There is no quality or size claim of ours on this panel, because we have not earned one. We did not compress this model, so any number here would be the publisher's, not ours. It is on the API because it is useful and cheap to run at volume, and it is in this list because you can call it, not because we changed it. The compression comparisons on this page are about Tileward 35B-A3B.
Your hardware, your VPC, the same governance.
66.9 GB → 24.5 GB
2.74× smaller than full precision, and the ratio matters more the larger the model gets. The expert layers, which are nearly all of the weight, go to 4 bits on a group of 32 rather than the usual 128. Attention, the router, the embedding and the output layer all stay at full precision, which is why the ratio is 2.74× and not higher.
OpenAI-compatible
Change the base URL and your existing client works. Same request shape hosted, in your VPC, or air-gapped.
Governed by the same tiles
The 214 Tileward Governance tiles apply to the self-hosted model exactly as they do to the hosted API, with audit records retained locally.
client.chat.completions.create(model="tileward-35b-a3b", messages=[...])
# same call self-hosted — only the base_url changes
Three ways to run it
- Change the base URL
- Default tile pack
- Zero infrastructure
- Single node, one GPU
- Custom tiles & policies
- Data never leaves
- Fully offline
- Flat licence, no metering
- Local audit retention