Tencent Hy3 Goes 1-Bit: Running a 295B Coding Model on a Single GPU

A 295B Model That Fits on One Card

Tencent's Hunyuan team released quantized versions of its flagship Hy3 model, a 295-billion-parameter MoE system. The 1-bit build (IQ1_M) compresses the weights from 598 GB down to 85.5 GiB — a 6.7x reduction — small enough to deploy on a single 96 GB inference GPU. A 4-bit build (Q4_K_M) comes in at 169.9 GiB and runs on two cards, staying close to full-model quality.

Tencent says the quantized versions hold up well on agent tasks, multilingual code, tool calling, and long-context understanding, with a GPTQ Int4 build available for vLLM and MTP speculative decoding to speed up the 1-bit variant. For a model this large, getting it onto a single GPU is a big deal — and it reopens a question every team eventually asks: is it cheaper to self-host or to pay per token?

The API Baseline

Hy3 is also available as a hosted API at roughly $0.20 input / $0.80 output per million tokens. That is already extremely cheap — cheaper than most Western frontier models by an order of magnitude. Any self-hosting decision has to beat that low bar, which changes the math considerably compared to self-hosting an expensive model.

The Self-Hosting Cost Math

Self-hosting the 1-bit Hy3 needs one 96 GB-class GPU. Whether you rent or buy, here is the shape of the cost:

• Rented cloud GPU: a 96 GB-class card runs roughly $2–$4 per hour, or about $1,400–$2,900 per month if you keep it running 24/7.
• Owned hardware: a suitable GPU is a five-figure purchase, amortized over a couple of years, plus power and maintenance.
• Utilization is everything. A rented GPU bills whether or not you send it work. If it sits idle half the day, your effective cost per token doubles.

Now the break-even. At $0.20/$0.80 per million tokens, a $1,400/month GPU rental only pays off if you would otherwise spend more than $1,400 on the API. Assuming a rough 1:1 input/output mix at a blended ~$0.50 per million, that is roughly 2.8 billion tokens per month before self-hosting the 1-bit model breaks even — and that is before you account for idle time, ops overhead, and the quality loss from 1-bit quantization.

When Self-Hosting Actually Wins

Because the Hy3 API is so cheap, the break-even volume is enormous. For most teams, the hosted API will simply be cheaper — you get frontier-scale capability at a fraction of a cent per thousand tokens with zero ops burden. Self-hosting the quantized model makes sense mainly when the driver is not raw price:

• Data residency and privacy — code that legally or contractually cannot leave your infrastructure.
• Extreme, sustained volume — a saturated GPU running near 100% utilization can beat even a cheap API at multi-billion-token scale.
• Latency control and offline operation — no dependence on an external provider's uptime or rate limits.

The real significance of the 1-bit release is not that everyone should now self-host. It is that the option is now cheap to explore: fitting a 295B model on one GPU turns a datacenter project into a single-machine experiment. That lowers the barrier to testing whether self-hosting fits your situation — even if, for most, the answer remains "just use the API."

Before you provision a GPU, run the numbers against the hosted rate. Estimate your monthly token volume in the AI Cost Estimator and compare it to the GPU rental cost — for cheap models like Hy3, the break-even is far higher than most teams expect.