GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
LessThanThreeAI
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
LessThanThreeAI
Model Tree
LessThanThreeAI/Qwen3.8-27B-Humanlike-Chat-2.0
Input Modalities
Output Modalities
Supported Functionality
Tested with vLLM 0.27.1 on an H100:
bash
vllm serve LessThanThreeAI/Qwen3.8-27B-Humanlike-Chat-2.0-FP8 --served-model-name humanlike-2.0 \ --max-model-len 16384 --language-model-only --linear-backend cutlass \ --enable-auto-tool-choice --tool-call-parser qwen3_coder --reasoning-parser qwen3On the H100, --linear-backend cutlass uses vLLM's precompiled block-FP8 kernel and the server is up in about 3 minutes. vLLM 0.27.1's default kernel on that card builds code at startup and needs nvcc in the container. On other GPUs, drop --linear-backend cutlass and vLLM picks a kernel for your card. Thinking is on by default. Send chat_template_kwargs: {"enable_thinking": false} per request for the fastest replies. Sampling: temperature 1.0, top_p 0.95, top_k 20.
Same layout as Qwen's official Qwen/Qwen3.8-27B-FP8: FP8 e4m3 weights in 128 x 128 blocks with BF16 scales, dynamic activation scaling. 407 linear layers are FP8: MLP, attention, the linear-attention projections and the MTP head's layers. These stay BF16: embeddings, lm_head, norms, the linear-attention gate tensors (in_proj_a, in_proj_b, A_log, dt_bias, conv) and the vision tower. Made from the BF16 weights with no calibration data.
The weights take 27.6 GiB in vLLM. A 48 GB card fits long contexts. A 32 GB card fits shorter contexts; lower --max-model-len and raise --gpu-memory-utilization to 0.95 if startup reports no memory for the cache. For 24 GB, use IQ4_XS or Q4_K_M from the main card.
Same check as every file on the main card: 5 fresh test chats against the reference (the base plus the 2.0 adapter at runtime in BF16), served on vLLM 0.27.1, one H100 NVL.
| File | KL vs reference | Top-1 agreement | Greedy turns identical | Check fails (greedy / sampled) | Decode tok/s |
|---|---|---|---|---|---|
| BF16 | 0.0032 | 97.2% | 12 of 15 | 0 / 0 | 46 |
| FP8 | 0.0079 | 91.7% | 9 of 15 | 0 / 0 | 72 |
Live requests on the same server: "yo you up" with thinking off came back yeah what's up with no reasoning. With thinking on, the train question came back 17:35 with the reasoning in its own field. A weather request produced get_weather with {"city": "Lisbon", "unit": "celsius"}.
Commissions open, DM codebottle on Discord (https://discord.com/users/320486798859960322). I build custom finetunes like this one: characters, product voices, distillation into smaller models, and domain or use-case specific models.
Apache-2.0, inherited from the upstream Qwen and Huihui releases.
vllm serve LessThanThreeAI/Qwen3.8-27B-Humanlike-Chat-2.0-FP8 --served-model-name humanlike-2.0 \ --max-model-len 16384 --language-model-only --linear-backend cutlass \ --enable-auto-tool-choice --tool-call-parser qwen3_coder --reasoning-parser qwen3