Tmax 27B — Hardware Requirements & GPU Compatibility
ChatFunctions
Tmax 27B is Allen AI's largest terminal-agent model, fine-tuned from Qwen 3.6 27B using DPPO, a reinforcement-learning method, to operate as a command-line coding agent rather than a general chat assistant. It is part of a family of terminal agents at 2B, 4B, 9B, and 27B scale, and the vision head from the base model was removed during training since it only handles text. On Terminal Bench 2.0 the card reports roughly 43% pass rate after 160 steps of RL training, ahead of the smaller Tmax variants. At 27B parameters, it needs a multi-GPU setup for full-precision inference, though it fits a single high-end GPU once quantized. The base architecture supports a context length of 262,144 tokens, though training used rollouts capped at 65,536 tokens per episode. It is released under the Apache 2.0 license, permitting unrestricted commercial and research use, and was published in June 2026.
Runs well on a 24 GB GPU
Runs from 16 GB (Q3_K_M · 32K)Best on 24 GB (Q4_K_M · ~39 tok/s · 128K)Macs from 18 GB (Q3_K_M · ~8 tok/s)
Apache 2.0
Reference GPU · RTX 4090 24 GB
Estimates, ±20% · how we estimate
Run Tmax 27B
One recommended way. Every other method, OS and quant is in the step-by-step guide.
Recommended
vllm serve allenai/tmax-27b --max-model-len 16384vLLM repo with its own quantisation
Download: Weights on Hugging Face
Other ways
each opens step-by-stepRun in cloud
Fits on RTX A5000 (24 GB) (6 GB headroom) · Q4_K_M
- Generation speed
- ~30 tok/s
- generation speed
- Cost per 1M output tokens
- $1.50
- per 1M output tokens
How Much VRAM Does Tmax 27B Need?
Select a quantization level; the listed hardware adheres to it.
| Quantization | VRAM | +32K | On RTX 4090 (reference) | |
|---|---|---|---|---|
| 12.2 GB | 13.5 GB | 2-bit · K-quant — 2-bit quantization with K-quant improvements | Fits: 128K context · ~54 tok/s | |
| 13.9 GB | 15.2 GB | 3-bit medium — 3-bit medium quantization | Fits: 128K context · ~47 tok/s | |
| 16.9 GB | 18.2 GB | 4-bit medium — 4-bit medium quantization — most popular sweet spot | Fits: 128K context · ~39 tok/s | |
| 19.9 GB | 21.3 GB | 5-bit medium — 5-bit medium quantization — good quality/size tradeoff | Fits: 64K context · ~33 tok/s | |
| 22.9 GB | 24.3 GB | 6-bit, very good — 6-bit quantization, very good quality | Fits: 16K context · ~29 tok/s | |
| 27.6 GB | 29 GB | 8-bit, near-lossless — 8-bit quantization, near-lossless | Does not fit: needs 28.1 GB | |
| 54.5 GB | 55.9 GB | 16-bit · bfloat — Brain floating point 16 — preferred for training | Does not fit: needs 55 GB |
Selected Q4_K_M
4-bit medium quantization17.3GBat 16K context
- Weights
- 16.1 GB
- KV cache
- 0.9 GB
- Overhead
- 0.3 GB
Fits RTX 4090 (reference) · 6.7 GB free · ~39 tok/s
≥50 tok/s needs ≥1.30 TB/s · RTX 5090 ~69
Which Hardware Can Run Tmax 27B?
Q4_K_M · 16.9 GBTmax 27B (Q4_K_M) requires 16.9 GB of VRAM to load the model weights. For comfortable inference with headroom for KV cache and system overhead, 22+ GB is recommended. Using the full 262K context window can add up to 13.9 GB, bringing total usage to 30.8 GB. 8 GPUs can run it, including NVIDIA GeForce RTX 5090, NVIDIA GeForce RTX 3090 Ti. 122 devices with unified memory can run Tmax 27B, including Mac Studio (M3 Ultra, 256GB), Mac Studio (M3 Ultra, 512GB), Mac Mini M5 Pro (24 GB).
20 GB
fits · 2.7 GB free · up to 32K ctx
1 GPU
RX 7900 XT~28Simulate output on RX 7900 XT32–40 GB
fits · 14.7 GB free · up to 256K ctx
1 GPU · 3 Mac classes · 4 device families
RTX 5090~69Simulate output on RTX 5090M5 Max36 GB~19Simulate output on Mac Studio M5 Max (36 GB)M532 GB~6Simulate output on MacBook Air 13" M5 (32 GB)Orin~8Simulate output on NVIDIA Jetson AGX Orin 32GBSnapdragon X Elite (X1E-80-100)~5Simulate output on Snapdragon X Elite Copilot+ PC
By VRAM tier · typical GPUs and Macs at Q4_K_M
Similar models
6| Model | Score | VRAM | tok/s | Why it is similar |
|---|---|---|---|---|
| Qwen3.8 27BAlibaba | 122.1 | 17.9 GB | ~38 | same size |
| Qwen3.6 27BAlibaba | 111.2 | 17.9 GB | ~38 | same size |
| Qwen3.5 27BAlibaba | 103.8 | 17.9 GB | ~38 | same size |
| Gemma 3 27B ITGoogle | 79.4 | 21.7 GB | ~36 | same size |
| Gemma 4 26B A4BGoogle | — | 16.9 GB | ~153 | same size |
| Ariarhymes-ai | — | 20.2 GB | ~168 | same size |
Specifications
- Publisher
- Allen AI
- Family
- Tmax
- Parameters
- 26.9B
- Architecture
- Qwen3_5ForConditionalGeneration
- Context Length
- 262,144 tokens
- Vocabulary Size
- 248,320
- Release Date
- 2026-06-21
- License
- Apache 2.0
Frequently Asked Questions
How much VRAM does Tmax 27B need?
Tmax 27B requires 16.9 GB of VRAM at Q4_K_M, or 54.5 GB at BF16. Full 262K context adds up to 13.9 GB (30.8 GB total).
VRAM = Weights + KV Cache + Overhead
Weights = 26.9B × 4.8 bits ÷ 8 = 16.1 GB
KV Cache + Overhead ≈ 0.8 GB (at 2K context + ~0.3 GB framework)
Fit ratings and hardware model lists check this model with room for a 16K-token context, which needs a little more memory.
KV Cache + Overhead ≈ 14.7 GB (at full 262K context)
VRAM usage by quantization
Can NVIDIA GeForce RTX 4090 run Tmax 27B?
Yes, at Q6_K (22.9 GB) or lower. Higher quantizations like Q8_0 (27.6 GB) exceed the NVIDIA GeForce RTX 4090's 24 GB.
What's the best quantization for Tmax 27B?
For Tmax 27B, Q4_K_M (16.9 GB) offers the best balance of quality and VRAM usage. Q5_K_M (19.9 GB) provides better quality if you have the VRAM. The smallest option is Q2_K at 12.2 GB.
VRAM requirement by quantization
★ Recommended — best balance of quality and VRAM usage.
Can I run Tmax 27B on a Mac?
Yes — MacBook Pro 14" M3 Pro (18 GB) and 108 other Macs can run Tmax 27B, the smallest at Q3_K_M. Apple Silicon uses unified memory, so the model shares RAM with the system. At Q4_K_M you need at least 16.9 GB of usable unified memory (RAM minus macOS overhead), from Mac Mini M5 Pro (24 GB).
Can I run Tmax 27B locally?
Yes — Tmax 27B can run locally on consumer hardware. Popular tools include Ollama, LM Studio, and llama.cpp.
How fast is Tmax 27B?
At Q4_K_M, Tmax 27B can reach ~284 tok/s on AMD Instinct MI350X. On NVIDIA GeForce RTX 4090: ~39 tok/s. Speed depends mainly on GPU memory bandwidth. Real-world results typically within ±20%.
tok/s = (bandwidth GB/s ÷ model GB) × efficiency
Example: NVIDIA B200 → 8000 ÷ 16.9 × 0.65 = ~308 tok/s
Estimated speed at Q4_K_M (16.9 GB)
Real-world results typically within ±20%. Speed depends on batch size, quantization kernel, and software stack.
What's the download size of Tmax 27B?
At Q4_K_M, the download is about 16.14 GB. The full-precision BF16 version is 53.79 GB. The smallest option (Q2_K) is 11.43 GB.
Which GPUs can run Tmax 27B?
8 consumer GPUs can run Tmax 27B at Q4_K_M. Top options include NVIDIA GeForce RTX 5090, AMD Radeon RX 7900 XT, AMD Radeon RX 7900 XTX. 1 GPU has plenty of headroom for comfortable inference.
Which devices can run Tmax 27B?
122 devices with unified memory can run Tmax 27B at Q4_K_M (16.9 GB), including AMD Ryzen AI 9 HX 370 (Strix Point) Laptop, ASUS Ascent GX10, Beelink GTR9 Pro (Ryzen AI Max+ 395, 128 GB), Framework Desktop (Ryzen AI Max+ 395, 128 GB). Apple Silicon Macs use unified memory shared between CPU and GPU, making them well-suited for local LLM inference.