Model Board 2026-08-22

The open weights we would put in your office, and the ones we would not.

Three models we deploy today, each with the job it wins at and the smallest Host that serves 50 seats. Below them, the models we track and will not sell, with the reason. Every figure here comes from the same sizing engine as your Estimate.

What we deploy

The Host column is the cheapest machine on AWS eu-central-1 that holds the model and stays above 22 tok/s per person using it at once, at 50 seats.

model	licence	parameters	wins at	context	smallest Host for 50 seats	expected tok/s
Qwen3-30B-A3B	Apache 2.0	30.5B total, 3.3B active (MoE)	everyday staff assistant, drafting	32k configuredup to 131k	1x L4 24GBg6.xlarge, AWQ 4-bit	265at 4 concurrent
Qwen3-32B	Apache 2.0	32.8B total (dense)	coding help	64k configuredup to 131k	1x L40S 48GBg6e.xlarge, AWQ 4-bit	300at 4 concurrent
gpt-oss-120b	Apache 2.0	116.8B total, 5.1B active (MoE)	the quality upgrade, on an 80GB card	128k configuredup to 128k	1x A100 80GB or 1x H100 80GBThis is the upgrade, not the default. We do not put it on a Plan unless you ask for it.	1062at 4 concurrent

Qwen3-30B-A3B

Only 3.3B parameters are active per token, so it answers fast on a small GPU and serves a whole company from one machine.

Size this model

Qwen3-32B

A dense model holds long code context together better than a sparse one, which is what your engineers will notice first.

Size this model

gpt-oss-120b

The strongest model that still fits one 80GB card. We do not put it on a default Plan, because the card it needs costs three to seven times the L4 that serves the same company well.

Ask us to price this

We have not benchmarked these ourselves

Every tok/s figure on this page is expected, derived from the published model specification and the published memory bandwidth of the card. We did not measure it, and we will not describe it as measured.

We replace each row with a measured figure from the first Deployment on that Host, taken from that customer’s Run Log, and we date-stamp the change. Until a row says measured, read it as arithmetic.

Cloud prices are published on-demand list rates gathered 2026-08-22. You pay them to your own provider, never to us.

board date: Model Board 2026-08-22
throughput basis: expected, from model and card specs
measured rows: 0 of 3

The honest part: models we track and will not sell you

These are good models. Some are better than the three above. None of them fits one machine we would hand over, so they are not on a Plan, and the reason is printed beside each one rather than left out.

model	licence	parameters	context	wins at	why it is not on a Plan
DeepSeek-V3.2	MIT	675B total (MoE)	164k	general reasoning, agentic work	Needs 8 GPUs to serve at 4-bit. There is no one-Host version of this, so we do not sell it.
GLM-4.6	MIT	355B total (MoE)	200k	very long documents, coding	Around 200GB of weights at 4-bit. Excellent, and out of reach of a single machine.
Kimi-K2	Modified MIT	1T total, 32B active (MoE)	256k	coding, agentic work	A trillion parameters does not fit one Host at any quantisation we would trust.
Qwen3-235B-A22B	Apache 2.0	235B total, 22B active (MoE)	128k	general reasoning	About 120GB at 4-bit, so two 80GB cards minimum. On the list for when we serve multi-GPU Hosts.
Llama-4-Scout-17B-16E-Instruct	Llama 4 Community Licence	109B total, 17B active (MoE)	10000k	long-context retrieval	Fits a large Host, but the licence carries naming and acceptable-use conditions we will not sign a customer up to by default.

When we serve multi-GPU Hosts, some of these move up. The Model Board email is where we say so, and we say it the month it happens.

Turn a row into a number

The Estimate takes the model above, your headcount, your cloud and your region, and gives you the monthly Host cost and the crossover against your seat bill. No signup, nothing to submit.

Size Qwen3-30B-A3B Size Qwen3-32B

Both links open with 50 seats on AWS eu-central-1 already answered. Move the slider and the arithmetic moves with it.

monthly, by email