MTPLX/Models/Inkling 2 Mini

The fastest way to run Inkling 2 Mini on a Mac.

MTPLX runs Inkling 2 Mini at up to 128.8 tok/s on a MacBook Pro M5 Max (Bare Speed, its fastest recorded answer, the model's own sampler). Thinking Machines Lab's open-weights coding and vision model runs on M3, M4 and M5 Macs from 24 GB and on M1 and M2 Macs from 32 GB with the FP16 builds, reads screenshots, and drives OpenCode, Pi and Hermes. MTPLX 2.13 adds it with Metal kernels written for this model.

The model

Inkling 2 Mini (Inkling-2-Mini on Hugging Face) is Thinking Machines Lab's open-weights model for coding, agents and images, released on 20 November 2026 under the Apache License 2.0. Its model card gives 26.4 billion parameters with 3.4 billion active for each token, a mixture of experts in which each token uses 8 of 128 experts, 12 middle layers that run three times, and a context window of up to 1M tokens. It reads images natively.

MTPLX 2.13 runs it on Apple Silicon with Metal kernels written for this model, and it uses the model's own multi-token-prediction draft head: the draft head proposes several tokens, the full model checks them all in one pass, and every kept token is a sample from the model's own distribution. There are three packs, each with an FP16 build for M1 and M2 Macs, and each carries the whole vision tower. Every pack needs MTPLX 2.13 or later.

Measured speeds

MacBook Pro M5 Max with 128 GB, MTPLX 2.13. Decode speed is the fastest recorded figure for each pack, at the model's own sampler: for Bare Speed, a 128-token answer to a 4,142-token prompt already in the prompt cache, recorded through mtplx serve on a 2.13 test build; for Optimized Speed, a 166-token answer in the 2.13 app; for Optimized Quality, the fastest stretch of an answer in the 2.13 app, its first 32 tokens. macOS thermal pressure was nominal for Bare Speed and Optimized Speed, and moderate when Optimized Quality's answer began. First token and prefill: mtplx serve, temperature 1.0 (the model's own sampler), fans at maximum, macOS thermal pressure nominal, on a new 4,142-token coding prompt, three runs for Bare Speed, one run for Optimized Speed on the release build and two runs for Optimized Quality (the mean where there is more than one), timed at the client: the first token is the time from sending the request to receiving it, and prefill is the prompt's tokens over that time. Long sessions on a hot Mac run slower than the peak.

PackDecode speed, peakFirst token, new promptPrefillDownload
Bare Speed128.8 tok/s1.69 s2,466 tok/s16.0 GB
Optimized Speed123.8 tok/s1.73 s2,397 tok/s17.2 GB
Optimized Quality103.8 tok/s1.84 s2,257 tok/s20.2 GB

Decode speed is each pack's fastest recorded figure, as the caption above says, and the first token is measured on a new 4,142-token prompt. Bare Speed is the first of the three that MTPLX lists. The FP16 builds have not run on an M1 or M2 yet, so no speed is listed for them. The same rows, with the quality checks, are on the benchmarks page.

The fastest way to run Inkling 2 Mini on a Mac

MTPLX is the fastest way to run Inkling 2 Mini on a Mac. On the same MacBook Pro M5 Max with a new 4,142-token coding prompt, Inkling 2 Mini Bare Speed on MTPLX shows its first token in 1.69 s against 2.83 s for oMLX and 2.26 s for plain mlx-lm, reads the prompt at 2,466 tok/s against 1,460 and 1,829, and then decodes 1.6 times as fast as oMLX and 2.8 times as fast as plain mlx-lm.

Engine and packFirst tokenPrefill
MTPLX, Inkling 2 Mini Bare Speed1.69 s2,466 tok/s
MTPLX, Inkling 2 Mini Optimized Speed1.73 s2,397 tok/s
MTPLX, Inkling 2 Mini Optimized Quality1.84 s2,257 tok/s
oMLX with its own draft head, official weights2.83 s1,460 tok/s
plain mlx-lm, official weights, no draft2.26 s1,829 tok/s

Measured on one MacBook Pro M5 Max with 128 GB on 22 November 2026, each engine through its own server with a fresh server for each run, the same 4,142-token prompt, temperature 1.0, top-p 1.0, top-k off and reasoning effort medium, three seeds (29, 31 and 37) taken in turn across the engines, every run starting and ending at macOS thermal pressure nominal. First token and prefill (the prompt's tokens over the first-token time) are timed at the client on the new prompt, and the decode ratio compares each engine's mean decode on a second request with that prompt cached; each figure is the mean over the runs (Inkling 2 Mini Optimized Quality: seeds 29 and 31, as the other run did not meet these conditions); Inkling 2 Mini Optimized Speed has no run on that sheet and is shown from one run on the release build. MTPLX 2.13.0. oMLX 0.8.0 (commit a9e23054) with its own draft head, and plain mlx-lm (commit 0c4c9919) with no draft, both on the official thinkingmachines/Inkling-2-Mini weights (revision 6e7c8248) with MLX 0.32.4.

Which pack for my Mac

MTPLX chooses for you. The app and mtplx start read your chip and memory, offer only the packs that fit, and size the context to your Mac. This is what MTPLX 2.13 offers for Inkling 2 Mini:

Your MacInkling 2 Mini pack to useAlso fitsContext
8 GB, M3, M4 or M5None fits. MTPLX recommends Qwen 3.5 4B Optimized Speed.
16 GB to 18 GB, M3, M4 or M5None fits. MTPLX recommends Bonsai 2 27B Optimized Speed.
24 GB, M3, M4 or M5Inkling 2 Mini Bare Speednone24,576 tokens
32 GB, M3, M4 or M5Inkling 2 Mini Bare SpeedOptimized Speed, Optimized Quality163,840 tokens
36 GB, M3, M4 or M5Inkling 2 Mini Bare SpeedOptimized Speed, Optimized Quality245,760 tokens
48 GB, M3, M4 or M5Inkling 2 Mini Bare SpeedOptimized Speed, Optimized Quality434,176 tokens
64 GB, M3, M4 or M5Inkling 2 Mini Bare SpeedOptimized Speed, Optimized Quality749,568 tokens
96 GB to 512 GB, M3, M4 or M5Inkling 2 Mini Bare SpeedOptimized Speed, Optimized Quality1,048,576 tokens
8 GB, M1 or M2None fits.
16 GB to 24 GB, M1 or M2None fits. MTPLX recommends Qwen 3.5 9B Optimized Speed FP16.
32 GB, M1 or M2Inkling 2 Mini Bare Speed FP16Optimized Speed FP16, Optimized Quality FP16163,840 tokens
64 GB, M1 or M2Inkling 2 Mini Bare Speed FP16Optimized Speed FP16, Optimized Quality FP16749,568 tokens
96 GB to 192 GB, M1 or M2Inkling 2 Mini Bare Speed FP16Optimized Speed FP16, Optimized Quality FP161,048,576 tokens
  • 24 GB: Bare Speed fits, with a 24,576-token context. On a 24 GB Mac the app lists Bonsai 2 27B first and Inkling 2 Mini Bare Speed second.
  • 32 GB and up: all three packs fit. Bare Speed comes first, Optimized Speed keeps 8-bit attention, and Optimized Quality is the official weights unchanged.
  • M1 and M2: the FP16 builds, from 32 GB. M1 and M2 have no native bfloat16, so these builds store the 16-bit weights as float16 and keep every 4-bit and 8-bit code the same. MTPLX picks them automatically. Their speed on an M1 or M2 has not been measured yet.
  • M3 and M4: they take a code path without the M5's tensor units. On an M5 Max forced onto that path, Bare Speed generated 106.1 tok/s on a cached conversation and answered a follow-up's first token in 0.06 s. An M3 or M4 has less memory bandwidth than an M5 Max, so expect lower figures there.
  • Below 24 GB: no Inkling 2 Mini pack fits. MTPLX recommends Qwen 3.5 4B Optimized Speed (8 GB) or Bonsai 2 27B Optimized Speed (16 GB to 18 GB) there instead.

The context column is the longest conversation MTPLX admits for the first pack on that much memory; the model itself goes up to 1,048,576 tokens.

Quality and checks

Each pack against the official Inkling 2 Mini weights, and the release checks every pack passed through mtplx serve at temperature 1.0:

PackSame top token as the official weightsLoss against the official weightsCoding testsScreenshot of a failing testTool call
Bare Speed239 of 256+0.0109 nats a token6 of 6read itcalled the tool
Optimized Speed247 of 256-0.0007 nats a token6 of 6read itcalled the tool
Optimized Qualitythe referencethe reference6 of 6read itcalled the tool
Bare Speed FP16241 of 256not measured yet6 of 6read itcalled the tool
Optimized Speed FP16246 of 256not measured yet6 of 6read itcalled the tool
Optimized Quality FP16245 of 256-0.0028 nats a token6 of 6read itcalled the tool

Six coding tests is a small sample: three tasks, two seeds each, run to completion and executed. The loss column is the change in average loss on 6,140 tokens of real code and prose against the official weights; a value inside two standard errors of zero is within noise. For the FP16 builds, the checks and the top-token figures ran on the code path an M1 or M2 takes, on an M5 Max.

How to run it

Mac app: download MTPLX, open it and pick "Inkling 2 Mini Bare Speed" (or the pack the app suggests for your Mac). The app downloads the pack, starts it, and opens OpenCode, Pi or Hermes on it with one click.

Command line, with Homebrew or pip:

brew install youssofal/mtplx/mtplx        # or: python3 -m pip install mtplx
mtplx pull Youssofal/Inkling-2-Mini-MTPLX-Bare-Speed
mtplx start opencode --model Youssofal/Inkling-2-Mini-MTPLX-Bare-Speed

mtplx start also opens pi and hermes. For anything else that speaks the OpenAI or Anthropic API, run mtplx serve --model Youssofal/Inkling-2-Mini-MTPLX-Bare-Speed and point it at http://127.0.0.1:8000. Sampling follows the model's own settings, temperature 1.0 over the whole distribution, unless the client sends its own.

Step by step, with the clients: Run Inkling 2 Mini on a Mac.

The six packs

PackForBuilt asDownload
Inkling 2 Mini Bare SpeedM3, M4 and M5 Macs with 24 GB or more4-bit attention, output head and embedding; official 4-bit experts16.0 GB
Inkling 2 Mini Optimized SpeedM3, M4 and M5 Macs with 32 GB or more8-bit attention and output head; original embedding; official 4-bit experts17.2 GB
Inkling 2 Mini Optimized QualityM3, M4 and M5 Macs with 32 GB or morethe official weights unchanged20.2 GB
Inkling 2 Mini Bare Speed FP16M1 and M2 Macs with 32 GB or moreBare Speed with its 16-bit parts in float1616.0 GB
Inkling 2 Mini Optimized Speed FP16M1 and M2 Macs with 32 GB or moreOptimized Speed with its 16-bit parts in float1617.2 GB
Inkling 2 Mini Optimized Quality FP16M1 and M2 Macs with 32 GB or moreOptimized Quality with its 16-bit parts in float1620.2 GB

How it is built

  • The experts stay as released. Every pack keeps the official 4-bit NVFP4 experts; Optimized Quality also keeps the official 8-bit MXFP8 experts in its last four layers.
  • Bare Speed stores the attention, the output head and the embedding in 4 bits (groups of 64). Optimized Speed stores the attention and the output head in 8 bits and keeps the original embedding. Optimized Quality changes nothing.
  • Vision and the draft head are whole. Every pack carries the model's whole vision tower and its draft head as released, in BF16, and in FP16 in the M1 and M2 builds.
  • Exact speculation. Drafts are accepted with the probability-ratio rule plus residual resampling, so the output follows the model's own distribution at any temperature.
  • Forge. mtplx forge inkling rebuilds each pack on your own Mac from the official weights, and mtplx forge fp16 makes the M1 and M2 builds.

Questions

What is the fastest way to run Inkling 2 Mini on a Mac?

MTPLX, the free Mac app and command line for local models on Apple Silicon. On a MacBook Pro M5 Max with 128 GB, MTPLX 2.13 runs Inkling 2 Mini Bare Speed at 128.8 tok/s (its fastest recorded answer, temperature 1.0, the model's own sampler), with the first token 1.69 s after a new 4,142-token prompt. It drafts with the model's own draft head and checks every token with the full model, so the answers keep the model's own distribution. Install the app from mtplx.com and pick Inkling 2 Mini Bare Speed, or run mtplx start opencode --model Youssofal/Inkling-2-Mini-MTPLX-Bare-Speed.

Which Inkling 2 Mini pack fits a 24 GB Mac?

Bare Speed. MTPLX 2.13 offers Inkling 2 Mini Bare Speed on M3, M4 and M5 Macs with 24 GB, with a 24,576-token context. Optimized Speed and Optimized Quality need 32 GB or more. On a 24 GB Mac the app lists Bonsai 2 27B first and Inkling 2 Mini Bare Speed second.

Does Inkling 2 Mini run on an M1 or M2?

Yes, on M1 and M2 Macs with 32 GB or more, through the FP16 builds: Inkling 2 Mini Bare Speed FP16, Optimized Speed FP16 and Optimized Quality FP16. M1 and M2 have no native bfloat16, so these builds store the 16-bit weights as float16 and keep every 4-bit and 8-bit code the same. MTPLX picks them automatically on those Macs. Their speed on an M1 or M2 has not been measured yet.

Can Inkling 2 Mini read images on a Mac?

Yes. Every MTPLX pack of Inkling 2 Mini carries the model's whole vision tower, so it reads screenshots, diagrams and photos in the app, Pi, OpenCode or any client that sends images. In MTPLX's release check each pack read a screenshot of a failing test and named the test and both counts.

How fast is Inkling 2 Mini on a Mac?

On a MacBook Pro M5 Max with 128 GB and MTPLX 2.13, Bare Speed, Optimized Speed and Optimized Quality decode at up to 128.8 tok/s, 123.8 tok/s and 103.8 tok/s, each pack's fastest recorded figure (temperature 1.0; Optimized Quality's is the first 32 tokens of an answer). The first token after a new 4,142-token prompt takes 1.69 s on Bare Speed. Long sessions on a hot Mac run slower than the peak.

How much memory does Inkling 2 Mini need on a Mac?

Bare Speed downloads 16.0 GB and peaked at 18.3 GiB serving a 4,142-token prompt; Optimized Speed downloads 17.2 GB (19.5 GiB) and Optimized Quality 20.2 GB (22.3 GiB). MTPLX offers Bare Speed from 24 GB and the other two from 32 GB, and sizes the context to your Mac: 163,840 tokens for Bare Speed on 32 GB, 749,568 on 64 GB, the full 1,048,576 from 96 GB.

Does Inkling 2 Mini work with OpenCode, Pi and Hermes?

Yes. mtplx start opencode, mtplx start pi and mtplx start hermes each start Inkling 2 Mini and open the client on it, with --model Youssofal/Inkling-2-Mini-MTPLX-Bare-Speed. The app does the same with one click, and mtplx serve gives any other client that speaks the OpenAI or Anthropic API the model on 127.0.0.1:8000.

Does MTPLX keep an Inkling 2 Mini coding session in the cache?

Yes. MTPLX keeps the conversation in memory between turns, so a follow-up turn reads only what is new. In sessions on MTPLX 2.13 with Inkling 2 Mini Bare Speed, the share of prompt tokens served from the cache after the first request was 99.7% in Pi (18 turns, to a 113,202-token context, with the server restarted in the middle), 91.8% in OpenCode (11 turns, with the server restarted in the middle), 97.9% in Hermes (11 turns, with the server restarted in the middle) and 89.8% in the app's Chat (10 messages, with the app quit and opened again in the middle). New text in a turn, such as a paste or a tool result, is read fresh, so no session reaches 100%. After a restart, Pi (1.45 s to the first token), Hermes (0.48 s to the first token) and the app's Chat (0.63 s to the first token) picked the conversation up from the SSD cache. OpenCode reads its conversation again after a restart (12.8 s to the first token for 22,070 tokens in this session), a known issue in 2.13.

Does MTPLX change Inkling 2 Mini's answers?

No. Drafts are accepted with the probability-ratio rule plus residual resampling, so the output follows the model's own distribution at any temperature. Optimized Quality is the official weights unchanged; Bare Speed picks the same top token as the official weights at 239 of 256 positions (93.4%) and Optimized Speed at 247 of 256 (96.5%), after a 4,142-token prompt.

Which Inkling 2 Mini pack should I pick?

Bare Speed is the first of the three that MTPLX lists and the smallest download. Optimized Speed stays closer to the original weights with 8-bit attention. Optimized Quality is the official release unchanged. MTPLX offers only the packs that fit your Mac.

Does Inkling 2 Mini work in MTPLX 2.12?

No. Inkling 2 Mini needs MTPLX 2.13 or later; MTPLX 2.12 and earlier cannot load it. The app updates itself; on the command line run brew upgrade mtplx or python3 -m pip install -U mtplx.

Credit. Inkling 2 Mini is by Thinking Machines Lab, released under the Apache License 2.0 with its acceptable use policy. Base model: thinkingmachines/Inkling-2-Mini. The model facts on this page come from its model card.