Models tested on PS5

Models tested on PS5

Every model here ran on a PS5 with Linux (40 GPU compute units, GPU memory set to max) on the Linux edition, and was checked against ts5's CPU reference (all pass except two Gemma cases noted below). Speeds are tokens per second, one request at a time, with no helper: prompt is reading a 512-token prompt, writing is generating the answer. Context is what -c auto gives with the current defaults. All files are Q4_K_M unless named.

The terminal app downloads every model on this page for you: F2, pick it under ⇣ Download, Enter, or type /download qwen3-14b (the keys are in its command list). The desktop app's Models page downloads the checked models, and its Search Hugging Face finds others. The commands below install these exact files from a shell instead.

The models

Model	Good for	File size	Prompt	Writing	Context	Helper to pair with it
Qwen3 14B	Best all-rounder: writing, explaining, reasoning; can think first	9.0 GB	~490	41	32k	Qwen3 0.6B
Qwen3.5 9B (DeepSeek-V4-Pro heretic)	Very long documents (found 10/10 facts at 225k tokens); fast follow-ups on the same document. Community fine-tune, uncensored	5.6 GB	~826	67	256k	none
DeepSeek R1 Distill Qwen 14B (abliterated v2)	Math, logic, puzzles; thinks out loud before answering. Refusals removed	9.0 GB	~490	38-41	32k	Qwen3 0.6B
Qwen3 8B	Fast general chat, thinking, tools	5.0 GB	~850	72	32k	Qwen3 0.6B
Llama 3.1 8B Instruct	Fast general chat with a long memory (10/10 facts at 111k)	4.9 GB	~850	74	128k	Llama 3.2 1B
Qwen 2.5 Coder 7B Instruct	Coding, fastest	4.7 GB	~960	77	32k	Qwen3 0.6B
Qwen 2.5 Coder 14B Instruct	Coding, stronger	9.0 GB	not measured	~28-30 (long prompts)	32k	Qwen3 0.6B
Gemma 3 12B	General chat; weak on long documents past ~8k (4/10 at 53k). One known near-tie in its check	7.3 GB	~406	45	64k	Gemma 3 1B
Gemma 4 12B	General chat; 6/10 facts at 112k. Its English-text check is over the bar (being looked into)	7.7 GB	~394	43	128k	Gemma 3 1B

Which one: Qwen3 14B for the best answers, Qwen3.5 9B for very long documents, Llama 3.1 8B or Qwen3 8B when speed matters most, Qwen 2.5 Coder for code. Long-document scores count 10 facts hidden in a document. More numbers are in Benchmarks and Test results on PS5 Linux.

Helpers are small models of the same family that draft a few words ahead for the big one, which then checks them. That makes writing faster with exactly the same text. The helper in the column is the one tuning picked on the test PS5. It's built into ts5 for these models: each uses it without any tuning once the helper's file is in the same folder (the terminal app downloads it right after the model).

Helper	Used for	File size
Qwen3 0.6B	Helper (draft) for the Qwen3, Qwen 2.5 Coder and DeepSeek models	0.4 GB
Llama 3.2 1B Instruct	Helper for Llama 3.1 8B and Llama 3.2 3B	0.8 GB
Gemma 3 1B	Helper for Gemma 3 12B and Gemma 4 12B	0.8 GB
Gemma 3 1B (Q8_0)	The same helper at 8 bits, slightly more accurate, a little larger	1.1 GB

Install from a terminal

ts5 fetch downloads over HTTPS, continues a stopped download where it left off, and checks the file's SHA-256 before keeping it. Each link below is pinned to the exact upload that was tested.

Run the commands inside the models folder of the way you use TokenStation5:

You use	Models folder	Run the commands as
The desktop app, the terminal app, or ts5 yourself	~/.local/share/tokenstation/models	yourself
The TV app (tokenstation-tv) or the API-only service (tokenstation)	/var/lib/tokenstation/models	the service's user: put sudo -u tokenstation in front of each ts5 fetch

The packages put ts5 in /usr/bin; the portable install puts it in ~/.local/opt/tokenstation/bin. A 14B model is about 9 GB, so check there's room (df -h .).

For the desktop app:

mkdir -p ~/.local/share/tokenstation/models
cd ~/.local/share/tokenstation/models

For the services:

cd /var/lib/tokenstation/models
# then: sudo -u tokenstation ts5 fetch ...

Qwen3 14B (9.0 GB, bartowski/Qwen_Qwen3-14B-GGUF)

ts5 fetch https://huggingface.co/bartowski/Qwen_Qwen3-14B-GGUF/resolve/bd080f768a6401c2d5a7fa53a2e50cd8218a9ce2/Qwen_Qwen3-14B-Q4_K_M.gguf Qwen_Qwen3-14B-Q4_K_M.gguf --sha256 915913e22399475dbe6c968ac014d9f1fbe08975e489279aede9d5c7b2c98eb6

Qwen3.5 9B (DeepSeek-V4-Pro heretic) (5.6 GB, mradermacher/DeepSeek-V4-Pro-Qwen3.5-9B-UNCENSORED-heretic-GGUF)

ts5 fetch https://huggingface.co/mradermacher/DeepSeek-V4-Pro-Qwen3.5-9B-UNCENSORED-heretic-GGUF/resolve/3b6f11e221195ee3650d7ac0ffce6f9fe7275d8f/DeepSeek-V4-Pro-Qwen3.5-9B-UNCENSORED-heretic.Q4_K_M.gguf DeepSeek-V4-Pro-Qwen3.5-9B-UNCENSORED-heretic.Q4_K_M.gguf --sha256 a692de9520e4c0bb1512deb91e950f38e14854650aff43419112f87e6052c44b

DeepSeek R1 Distill Qwen 14B (abliterated v2) (9.0 GB, mradermacher/DeepSeek-R1-Distill-Qwen-14B-abliterated-v2-GGUF)

ts5 fetch https://huggingface.co/mradermacher/DeepSeek-R1-Distill-Qwen-14B-abliterated-v2-GGUF/resolve/39e97fd1a97e3e4b5be8ba92d6640a95f2757d29/DeepSeek-R1-Distill-Qwen-14B-abliterated-v2.Q4_K_M.gguf DeepSeek-R1-Distill-Qwen-14B-abliterated-v2.Q4_K_M.gguf --sha256 5bc6f9ce7fbc305c4eeb06253c251e69d9920a5fbac352d73654b63ea8d17e70

Qwen3 8B (5.0 GB, bartowski/Qwen_Qwen3-8B-GGUF)

ts5 fetch https://huggingface.co/bartowski/Qwen_Qwen3-8B-GGUF/resolve/0b69f75b7472688e6808490aa2b85efdb81b5ce7/Qwen_Qwen3-8B-Q4_K_M.gguf Qwen_Qwen3-8B-Q4_K_M.gguf --sha256 54fffa050078e984116639c83dfb64b5aa6d4cd474e018b076777c632bbccccd

Llama 3.1 8B Instruct (4.9 GB, bartowski/Meta-Llama-3.1-8B-Instruct-GGUF)

ts5 fetch https://huggingface.co/bartowski/Meta-Llama-3.1-8B-Instruct-GGUF/resolve/bf5b95e96dac0462e2a09145ec66cae9a3f12067/Meta-Llama-3.1-8B-Instruct-Q4_K_M.gguf Meta-Llama-3.1-8B-Instruct-Q4_K_M.gguf --sha256 7b064f5842bf9532c91456deda288a1b672397a54fa729aa665952863033557c

Qwen 2.5 Coder 7B Instruct (4.7 GB, bartowski/Qwen2.5-Coder-7B-Instruct-GGUF)

ts5 fetch https://huggingface.co/bartowski/Qwen2.5-Coder-7B-Instruct-GGUF/resolve/1f629da0c8bed16b9e50cee91c70693650e66c35/Qwen2.5-Coder-7B-Instruct-Q4_K_M.gguf Qwen2.5-Coder-7B-Instruct-Q4_K_M.gguf --sha256 1664fccab734674a50763490a8c6931b70e3f2f8ec10031b54806d30e5f956b6

Qwen 2.5 Coder 14B Instruct (9.0 GB, bartowski/Qwen2.5-Coder-14B-Instruct-GGUF)

ts5 fetch https://huggingface.co/bartowski/Qwen2.5-Coder-14B-Instruct-GGUF/resolve/5b379ec4bf71bafecb5f9081ad28b19939128988/Qwen2.5-Coder-14B-Instruct-Q4_K_M.gguf Qwen2.5-Coder-14B-Instruct-Q4_K_M.gguf --sha256 2946d28c9e1bb2bcae6d42e8678863a31775df6f740315c7d7e6d6b6411f5937

Gemma 3 12B (7.3 GB, bartowski/google_gemma-3-12b-it-GGUF)

ts5 fetch https://huggingface.co/bartowski/google_gemma-3-12b-it-GGUF/resolve/648e3a36a77c8a9f12d86e741f9dcb9089c769c4/google_gemma-3-12b-it-Q4_K_M.gguf google_gemma-3-12b-it-Q4_K_M.gguf --sha256 fc57f67efa46d711c346e587cbef7d049e95f3df8db2eb2271153343ef0acc7b

Gemma 4 12B (7.7 GB, bartowski/gemma-4-12B-it-GGUF)

ts5 fetch https://huggingface.co/bartowski/gemma-4-12B-it-GGUF/resolve/2ae7d41be21ca62de00a2d320ee9cec50daa3aa6/gemma-4-12B-it-Q4_K_M.gguf gemma-4-12B-it-Q4_K_M.gguf --sha256 3962624dcd25b947d889dc9ae1bf275b61db6cd4dbe694057f34fffef1671509

Helpers

Qwen3 0.6B (0.4 GB, unsloth/Qwen3-0.6B-GGUF)

ts5 fetch https://huggingface.co/unsloth/Qwen3-0.6B-GGUF/resolve/50968a4468ef4233ed78cd7c3de230dd1d61a56b/Qwen3-0.6B-Q4_K_M.gguf Qwen3-0.6B-Q4_K_M.gguf --sha256 ac2d97712095a558e31573f62f466a3f9d93990898b0ec79d7c974c1780d524a

Llama 3.2 1B Instruct (0.8 GB, bartowski/Llama-3.2-1B-Instruct-GGUF)

ts5 fetch https://huggingface.co/bartowski/Llama-3.2-1B-Instruct-GGUF/resolve/067b946cf014b7c697f3654f621d577a3e3afd1c/Llama-3.2-1B-Instruct-Q4_K_M.gguf Llama-3.2-1B-Instruct-Q4_K_M.gguf --sha256 6f85a640a97cf2bf5b8e764087b1e83da0fdb51d7c9fab7d0fece9385611df83

Gemma 3 1B (0.8 GB, bartowski/google_gemma-3-1b-it-GGUF)

ts5 fetch https://huggingface.co/bartowski/google_gemma-3-1b-it-GGUF/resolve/116f76234503685a98f572982177b11d44ec8ff1/google_gemma-3-1b-it-Q4_K_M.gguf google_gemma-3-1b-it-Q4_K_M.gguf --sha256 12bf0fff8815d5f73a3c9b586bd8fee8e7b248c935de70dec367679873d0f29d

Gemma 3 1B (Q8_0) (1.1 GB, bartowski/google_gemma-3-1b-it-GGUF)

ts5 fetch https://huggingface.co/bartowski/google_gemma-3-1b-it-GGUF/resolve/116f76234503685a98f572982177b11d44ec8ff1/google_gemma-3-1b-it-Q8_0.gguf google_gemma-3-1b-it-Q8_0.gguf --sha256 375e12a4a18929a641f9744b060d4a7cf4e279530750555828ec0c117870bc96

After downloading

• The app and the terminal app list a new model within a few seconds. The API-only service uses MODEL in /etc/default/tokenstation (or the first file in its folder); restart it after a change.
• These models need no tuning: their settings are built in, and each uses its helper when the helper's file is in the same folder. ts5-tune --models (or T on the terminal app's Models page) is for other models, or to measure again on your own machine.
• Each model is under its own license: Llama and Gemma have terms of use of their own. The model files are not part of TokenStation5.