Inkling Small 276B-A12B needs about 168 GB at 4-bit compression with an 8,192-token conversation (about 6,000 words): more than any PC holds in graphics memory. It needs a workstation or server with that much system memory, where it can still write at a usable speed: each token uses only 12 billion of its parameters. Of the PCs below, the least powerful that runs it well is an RTX 5070 Ti with 64 GB of DDR5: 21–33 tokens/s at 1.6-bit.
A second, different card shares the model with the first.
Choosing one fills in the memory it usually has.
Settings, System, About: “Installed RAM”.
Task Manager, Performance, Memory: “Speed”.
| Model | Runs? | Writes | Needs |
|---|
The model’s weights plus the conversation it is holding, at each compression. For it to run at full speed, all of it has to fit in the graphics card’s memory with about 1.2 GB to spare. Slow although it should fit? See where it is really running. Which compression to download.
| Compression | Weights | 8,192 tokens | 32,768 tokens | 131,072 tokens |
|---|---|---|---|---|
| 16-bit (BF16) | 552 GB | 552 GB | 553 GB | 556 GB |
| 8-bit (Q8_0) | 293 GB | 294 GB | 294 GB | 297 GB |
| 6-bit (Q6_K) | 226 GB | 227 GB | 227 GB | 230 GB |
| 5-bit (Q5_K_M) | 196 GB | 196 GB | 197 GB | 199 GB |
| 4-bit (Q4_K_M) | 167 GB | 168 GB | 168 GB | 171 GB |
| 3-bit (Q3_K_M) | 135 GB | 135 GB | 136 GB | 139 GB |
| 2-bit (Q2_K) | 104 GB | 104 GB | 105 GB | 107 GB |
| 2-bit (IQ2_XXS) | 76 GB | 76 GB | 77 GB | 80 GB |
| 1.6-bit (IQ1_S) | 55 GB | 56 GB | 56 GB | 59 GB |
With an 8,192-token conversation, at the best compression that fits each PC. Writing speed is how fast an answer appears; most people read at about 5 to 8 tokens a second.
| PC | Runs? | Writes | How |
|---|---|---|---|
| Raspberry Pi 5, 8 GB, microSD | Too slow to use | 0.01–0.01 tokens/s | 1.6-bit (IQ1_S), partly read from the drive as it runs |
| Mini PC, Intel N100, 16 GB DDR5-4800 (one channel) | Too slow to use | 0.30–0.48 tokens/s | 1.6-bit (IQ1_S), partly read from the drive as it runs |
| Laptop, Core i5-8250U, 8 GB DDR4-2400, SATA SSD | Too slow to use | 0.04–0.07 tokens/s | 1.6-bit (IQ1_S), partly read from the drive as it runs |
| Laptop, Ryzen 7 7840HS with Radeon 780M, 32 GB DDR5-5600 | Too slow to use | 0.91–1.5 tokens/s | 1.6-bit (IQ1_S), partly read from the drive as it runs |
| Laptop, Core Ultra 7 258V with Arc 140V, 32 GB LPDDR5X-8533 | Too slow to use | 0.91–1.4 tokens/s | 1.6-bit (IQ1_S), partly read from the drive as it runs |
| Desktop, no graphics card, 32 GB DDR5-5600 | Too slow to use | 0.90–1.4 tokens/s | 1.6-bit (IQ1_S), partly read from the drive as it runs |
| RTX 4060 8 GB, 32 GB DDR5-5600 | Too slow to use | 1.2–1.9 tokens/s | 1.6-bit (IQ1_S), partly read from the drive as it runs |
| RTX 3060 12 GB, 32 GB DDR4-3200 | Too slow to use | 1.4–2.2 tokens/s | 1.6-bit (IQ1_S), partly read from the drive as it runs |
| RTX 5060 Ti 16 GB, 32 GB DDR5-6000 | Too slow to use | 1.8–2.9 tokens/s | 1.6-bit (IQ1_S), partly read from the drive as it runs |
| RX 9070 XT 16 GB, 32 GB DDR5-6000 | Too slow to use | 1.8–2.9 tokens/s | 1.6-bit (IQ1_S), partly read from the drive as it runs |
| RTX 5070 Ti 16 GB, 64 GB DDR5-6000 | Runs fast | 21–33 tokens/s | 1.6-bit (IQ1_S), graphics card and system memory |
| RTX 3090 24 GB, 64 GB DDR4-3600 | Runs well | 10–16 tokens/s | 2-bit (IQ2_XXS), graphics card and system memory |
| RX 7900 XTX 24 GB, 64 GB DDR5-6000 | Runs well | 16–26 tokens/s | 2-bit (IQ2_XXS), graphics card and system memory |
| RTX 4090 24 GB, 64 GB DDR5-6000 | Runs well | 16–26 tokens/s | 2-bit (IQ2_XXS), graphics card and system memory |
| RTX 5090 32 GB, 96 GB DDR5-6000 | Runs well | 12–20 tokens/s | 2-bit (Q2_K), graphics card and system memory |
| RTX 5090 and RTX 3090 (56 GB), 128 GB DDR5-6000 | Runs well | 8.1–13 tokens/s | 4-bit (Q4_K_M), graphics card and system memory |
| Two RTX 3090 (48 GB), 128 GB DDR4-3200 | Runs well | 8.3–13 tokens/s | 2-bit (Q2_K), graphics card and system memory |
| Mac mini, M4 Pro, 64 GB | Runs fast | 38–61 tokens/s | 1.6-bit (IQ1_S), graphics and the rest of memory |
| Ryzen AI Max+ 395, 128 GB | Runs fast | 25–40 tokens/s | 2-bit (Q2_K), graphics and the rest of memory |
| NVIDIA DGX Spark, 128 GB | Runs fast | 32–51 tokens/s | 2-bit (Q2_K), in the graphics share of its memory |
| Mac Studio, M3 Ultra, 512 GB | Runs fast | 20–32 tokens/s | 8-bit (Q8_0), in the graphics share of its memory |
| RTX PRO 6000 96 GB, 512 GB DDR5-6400 in 8 channels | Runs fast | 39–63 tokens/s | 4-bit (Q4_K_M), graphics card and system memory |
| Server, EPYC 9755, no graphics card, 1,536 GB DDR5-6400 in 12 channels | Runs fast | 32–52 tokens/s | 4-bit (Q4_K_M), on the processor, in system memory |
| Server: RTX PRO 6000 96 GB, 1,536 GB DDR5-6400 in 12 channels | Runs fast | 51–81 tokens/s | 4-bit (Q4_K_M), graphics card and system memory |
| Data-centre server: eight H100 80 GB (640 GB) | Runs fast | 61–98 tokens/s | 8-bit (Q8_0), on the graphics card |
| Made by | Thinking Machines |
|---|---|
| Released | July 2026 |
| Parameters | 276 billion, 12 billion per token |
| Longest conversation | not stated |
| Good for | Chat and writing, images, reasoning and agents and tools |
| Licence | Apache 2.0 |
From its page on Hugging Face, read on 8 November 2026. LM Studio, Ollama and Jan load it from GGUF files such as unsloth/Inkling-Small-GGUF. Bottleneck is not connected with Thinking Machines.
Writing an answer is limited by memory rather than arithmetic: for every word-piece (token) a model writes, the computer reads every weight that token uses once. So the speed is how fast the memory holding the model can be read, divided by how much is read for each token: the model’s weights at the chosen compression, plus the conversation so far. How much of its memory speed each graphics card reaches is measured, not assumed: it comes from the card’s own result in llama.cpp’s benchmark, which LM Studio, Ollama and Jan are built on. A model that does not fit in the graphics card’s memory spills into system memory, five to fifteen times slower, and one that does not fit in memory at all can be read from the drive as it runs, slower again; the estimate counts each part at its own speed, and a processor with few threads cannot use all of fast memory. Graphics built into a processor use system memory, at its speed. Reading a prompt is limited by arithmetic, so it follows the measured prompt speed.
A model that is slower than it should be is nearly always partly on the processor rather than all on the graphics card, which is several times slower — and Ollama does that without a word when something else holds the card’s memory. Bottleneck’s free app checks a model run with Ollama, LM Studio, llama.cpp, KoboldCpp or Jan: where it really runs, how fast it writes, and what your PC should manage with it. Nothing leaves your PC. Or check and fix it by hand.
It reads your parts too: the graphics card and how much memory it really has free, how much system memory is installed, the speed it runs at and in how many channels. Memory left in one channel, or below its rated speed, halves how fast a model runs from it, and Bottleneck says so. Its AI models screen then lists what runs on this PC.
The free version shows where a model runs and how fast, with no time limit and no account. Pro, $9 once, explains why it is slower than it could be and how to fix it. Compare the versions