Tileward 35B-A3B
Compressed from Qwen3.6-35B-A3B. Serving requests today.
- Weights
- 68.6 GB → 24.5 GB, 64% smaller
- Active
- ~3B of 35B parameters per token
- Context
- 262,144 tokens
Tileward / Models
01 · Tileward Models
Governed · Compressed: run a bigger model in your VPC, on an air-gapped network, or on the GPUs you have.
A model that does not fit on the card you have costs a second card, a smaller model, or a cloud bill. Tileward Models fits it.
Works with the clients you already run
28 have a recipe to paste, and anything else that speaks MCP works too.
Quality
A compressed model has three jobs. Two are measured, one is where it runs.
Fit on the hardware you have, stay the model you chose, and run wherever you run it. The first two are measured against the publisher’s own release.
Fits the card you have
24.5 GB
of weights for Tileward 35B-A3B, from 68.6 GB as published: 2.8× smaller, and 1.46× smaller than the publisher’s own FP8 release.
Still the model you chose
±0.3
is the error bar on each score, and the gap between our build and the publisher’s FP8 release sits inside it: level on a multiple-choice test.
Runs where you run it
3
places it runs unchanged: our hosted API, your VPC, or air-gapped.
Catalogue
Every model we have compressed.
Three are served today and one is no longer served. The rest are compressed and ready to serve.
Tileward-Qwen3.8-27b
Compressed from Qwen3.8-27B. Serving requests today.
- API id
- Tileward-Qwen3.8-27b
- Input
- Text and images
gpt-oss 20B
Served as its publisher shipped it.
- Active
- ~3.6B of 20B parameters per token
- Context
- 8,192 tokens
- Locking
- Not available on this model
GLM-5.3-Flash
Compressed, not served. Hosted or on your hardware, on request.
Llama-3.1-Nemotron-Ultra-253B
Compressed, not served. Hosted or on your hardware, on request.
Nemotron 3 Nano 30B
Compressed, not served. Hosted or on your hardware, on request.
Qwen3.5-397B
Compressed, not served. Hosted or on your hardware, on request.
Qwen3-Next-80B-A3B
Compressed, not served. Hosted or on your hardware, on request.
Qwen3-Coder-30B-A3B
Compressed, not served. Hosted or on your hardware, on request.
gpt-oss 120B
No longer served.
Running one of the unserved ones? We can serve it for you, hosted or on your hardware.
Served builds
Every model we serve, spec by spec.
Weights, memory, context window and locking for each served build.
Tileward 35B-A3B: Qwen3.6-35B-A3B, compressed to just over a third
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..
| Measure | Result |
|---|---|
| Base | Qwen3.6-35B-A3B |
| Parameters | 35B total, ~3B active |
| Size on disk | Weights on disk, measured rather than estimated, on the catalog’s MiB/1000 basis. |
| Smaller by | 64%, a 2.8× ratio |
| Context window | 262,144 tokens |
| Per-tile locking | Available |
Measured against the publisher’s FP8 build: the eval · the write-up.
Tileward-Qwen3.8-27b: Qwen3.8-27B, compressed by Tileward
Serving requests today, with text and image input.
| Measure | Result |
|---|---|
| Base | Qwen3.8-27B |
| API id | Tileward-Qwen3.8-27b |
| Architecture | Dense |
| Input | Text and images |
| Compressed by Tileward | Yes |
gpt-oss 120B: no longer served
Third-party open weights under Apache 2.0 that we no longer serve.
| Measure | Result |
|---|---|
| Parameters | 117B total, ~5.1B active |
| Weights | As published, unchanged |
| Compressed by Tileward | No |
| On the API | No longer served |
gpt-oss 20B: served exactly as published
Third-party open weights under Apache 2.0, served as its publisher shipped 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 |
What you get
Your hardware, your VPC, the same governance.
1.8× to 2.8× smaller
Tileward 35B-A3B, a mixture-of-experts, lands 2.8× smaller than the weights it was compressed from. The dense Tileward-Qwen3.8-27b lands 1.8×, because more of a dense model stays at full precision.
OpenAI-compatible
Change the base URL and your existing client works.
Governed by the same tiles
The 215 Tileward Governance tiles apply to the self-hosted model exactly as they do to the hosted API, with audit records retained locally.
client = openai.OpenAI(base_url="https://api.tileward.com/v1", api_key=TILEWARD_KEY)
client.chat.completions.create(model="Tileward-Qwen3.6-35B-A3B", messages=[...])
# also answers to tileward-35b-a3b and Qwen3.6-35B-A3B-TW, at the same rate
# same call self-hosted, only the base_url changes
client.chat.completions.create(model="Tileward-Qwen3.6-35B-A3B", messages=[...])
# also answers to tileward-35b-a3b and Qwen3.6-35B-A3B-TW, at the same rate
# same call self-hosted, only the base_url changes
Three ways to run it
Hosted API
Minutes
- Change the base URL
- Default tile pack
- Zero infrastructure
Your VPC
Days
- Single node, one GPU
- Tiles and policies of your own
- Data never leaves
On-prem, air-gapped
Weeks
- Fully offline
- Runs on your hardware
- Local audit retention
Run it on your own hardware