docker run --rm --gpus all -p 8000:8000 docker.io/textclf/tq-quant:4bit-main vllm serve textclf/Qwen3-Coder-Next-TQ-4bit --quantization tqQuant Factory / early access
Quantize and fine-tune any model. Run it anywhere.
Quant Factory is the open-source, calibration-free toolkit for turning model weights into compact CQ models. Test the full workflow on your own GPU, then use our upcoming platform to fine-tune quantized models for enterprise workloads.
Free during early access — try it before it becomes paid
Now available in preview
Meet Quant Factory: our open-source, calibration-free path from model weights to CQ inference. Fine-tuning for quantized models is coming soon, with an enterprise platform designed for private workloads.
Try Quant Factory locally →The Quant Factory workflow
From weights to enterprise-ready models.
01
Start with weights
Bring a model from Hugging Face.
02
Quantize without calibration
No dataset. No calibration forward passes.
03
Fine-tune soon
Adapt quantized models for private enterprise workloads.
Open source tooling. CQ compression. Your GPU.
Quant Factory / early access
Quantize it. Run it. Decide for yourself.
Quant Factory is our open-source toolkit for turning Hugging Face models into ThousandML CQ 4-bit weights — without calibration data. Fine-tuning for quantized models is coming soon: we are building a secure platform for enterprises to adapt compact models to private data and production workloads.
Coming soon / enterprise fine-tuning
Use the open-source workflow today. Soon, enterprises will be able to fine-tune quantized models on private data through the ThousandML platform.
CQ 4-bit performance
Qwen3.8 27B, quantized without calibration
WikiText evaluation
92.4%
top-1 agreement with full precision
0.0282
mean KLD at CQ 4-bit
21.3 tok/s
measured throughput in production
No calibration dataset. No calibration forward passes. Quant Factory produces a compact CQ model directly from the weights, then lets you verify the result on your own GPU.
01 / quantize
./quantize-model meta-llama/Llama-3.1-8B-Instruct02 / build
python build_convert_upload_tq_model_incremental_qwen4exp.py --model meta-llama/Llama-3.1-8B-Instruct --repo-id YOUR_USERNAME/Llama-3.1-8B-Instruct-TQ-4bit --no-upload03 / install CQ support
cd inference
pip install .04 / serve
vllm serve YOUR_USERNAME/Llama-3.1-8B-Instruct-TQ-4bit \
--quantization tqBring your own model with Quant Factory, or start with one of the ready-to-run commands below. Either way, you get an OpenAI-compatible endpoint at localhost:8000 for testing with your own prompts and workloads.
docker run --rm --gpus all -p 8000:8000 docker.io/textclf/tq-quant:4bit-main vllm serve textclf/Qwen3.8-27B-TQ-4bit --quantization tqdocker run --rm --gpus all -p 8000:8000 docker.io/textclf/tq-quant:4bit-main vllm serve textclf/Llama-3.3-70B-Instruct-TQ-4bit --quantization tqdocker run --rm --gpus all -p 8000:8000 docker.io/textclf/tq-quant:4bit-main vllm serve textclf/Llama-3.1-8B-Instruct-TQ-4bit --quantization tqThese are the same CQ quantized weights used behind our hosted inference. Run them on your own GPU today; enterprise fine-tuning for quantized models is coming soon through the ThousandML platform.
Requires Docker, a CUDA-capable GPU, and a local Hugging Face cache. Each command serves an OpenAI-compatible endpoint on port 8000.
The technology
The same models, for a fraction of the price.
To run big models cheaply, providers quantize them — and most quantize so aggressively that they quietly trade away accuracy. Our ThousandML CQ doesn't: it makes models far cheaper to run while keeping their answers just as good — and we hand every bit of that saving straight back to you as a lower price per token.
4.1x
smaller footprint
99.6%
accuracy retained vs. FP16
CQ 4-bit · Qwen 3.8 27B · WikiText
0.0282
mean KL divergence
92.4%
top-1 agreement
On par with leading quants — but achieved without any calibration data, so the quality carries over to tasks well beyond the benchmark.
The most accuracy per bit
At any given size — say 4 bits per weight — there’s a mathematical best for how much accuracy you can keep. Most methods leave a lot on the table; our ThousandML CQ gets remarkably close to that best, so at the same size our models stay noticeably sharper.
Data-free, so it holds up everywhere
Most methods tune their quant on a calibration dataset, which quietly overfits to whatever it was tuned on — great on that test, shakier everywhere else. Ours uses no calibration data at all, so the quality holds steady across the full range of tasks, not just the ones it was measured on.
Same answers, proven
We test every model against the untouched original and keep it within a fraction of a percent on standard benchmarks — no silent quality drops.
The savings go straight to you
Smaller models are cheaper for us to run — and we don’t pocket the difference. Every bit of efficiency ThousandML CQ unlocks is passed directly to you as a lower price per token.
One API. Every model worth running.
Switch models with a single string. Every model on ThousandML is served from the same OpenAI-compatible endpoint at the same low, transparent per-token price.
ModelProviderInputOutputCached inputThroughput
Llama 3.3 70B Instruct
128K context
Meta
Input $0.08Output $0.30Cached input —6.6 tok/sLlama 3.1 8B Instruct
128K context
Meta
Input $0.018Output $0.038Cached input —21.2 tok/sQwen3.8 27B
262K context
Alibaba
Input $0.30Output $2.50Cached input $0.0321.3 tok/sQwen3 Coder Next
262K context
Alibaba
Input $0.10Output $0.78Cached input $0.03—DeepSeek V4 FlashSoon
1M context
DeepSeek
Input $0.05Output $0.12Cached input $0.01—MiMo V2.5Soon
— context
Xiaomi MiMo
Input $0.11Output $0.23Cached input $0.00255—
Prices shown per 1M tokens (USD). 100+ additional models available.
View full model catalog →Change one line. Cut your bill.
Already using the OpenAI SDK? Point the base URL at ThousandML and keep the rest of your code exactly as it is.
- Drop-in replacement for the OpenAI SDK
- Same request and response schema
- Streaming, tools, and JSON mode supported
quickstart.sh
curl https://api.textclf.com/v1/chat/completions \
-H "Authorization: Bearer YOUR_TEXTCLF_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "Qwen/Qwen3.8-27B",
"messages": [
{ "role": "user", "content": "Explain inference cost in one line." }
]
}'Prepaid credit. Zero surprises.
Get $5 in free credit to start — no credit card needed. When you're ready for more, top up your balance and spend it at each model's per-token rate. You can never be charged more than you've added.
Free credit
$5on us
Start building the moment you sign up — no credit card required.
- $5 in credit, free
- No credit card to start
- Access to every model
- OpenAI-compatible API
Prepaid credits
Most popularTop upfrom $10
Add credit and draw it down as you call models. Each model bills at its own per-token rate — you only spend what you use.
- Per-token pricing, set per model
- Credits never expire
- Auto-reload so you never run dry
- Real-time usage & spend dashboard
Enterprise
Customvolume
Committed-use rates, dedicated capacity, and invoicing.
- Volume discounts on credit
- Dedicated GPU capacity
- SOC 2 Type II, SSO/SAML
- Invoicing & dedicated support
Every model has its own input and output rate. See per-token pricing for all models.