Model Follower.

Up to date

Rebuilt
2026-10-25 06:54 UTC
Release radar
117 on watch
Hub snapshot
2026-10-25
Re-checked
daily · every number sourced and dated

Models › Inkling-Small family

Inkling-Small

Hugging Face ↗ Follow changes

thinkingmachines Apache 2.0 out 3 months ago

266B parameters (6 of 256 experts per token), images and text in, text out, 1M tokens of context. 2 providers serve it from $0.45 per million input tokens.

Verified 2026-10-25 Rank #329 in LLMs · score 82.1 method v1 Sources 4 Downloads 30d 989.9k · #110 of 619 tracked

Changed this week: 2 new quantized builds on the Hub (56 in all).2026-10-24

Can I run it?160GB at INT4Estimatefits 2×80 GBmethod
What does it cost?$0.45/1M incheapest of 1DeepInfrafp82026-10-25
Can I use it?Apache 2.0Permissive; patent grantlicence file2026-10-24
Do people use it?989.9kdownloads, 30 days#110 of 619 we track · top 18%+8 likes, downloads +28% this week429 likesHugging Face2026-10-24

Worth reading

Picked by rule from 3 tracked pieces: the lab's own post, the most shared, a field report, the latest.

Inkling-Small surpasses Inkling on reasoning and agentic coding benchmarks despite Inkling maintaining an advantage on knowledge coverage and factuality.

From Inkling-Small · thinkingmachines.ai · 2026-07-30
Most discussed · 33 HN pts

Inkling-Small

Thinking Machines releases Inkling-Small, an efficient open-weights Mixture-of-Experts model with 276B total parameters and 12B active trained on NVIDIA GB300 NVL72 systems.

  • Inkling-Small surpasses Inkling on reasoning and agentic coding benchmarks despite Inkling maintaining an advantage on knowledge coverage and factuality.
  • The model uses an encoder-free architecture where images are divided into 40×40-pixel patches and transformed using a four-layer hMLP.
thinkingmachines.ai 2026-07-30 33 HN pts · 2 comments 2.4k words sells inference

3 best of 3 tracked

The 2 picks above, plus these 1. Top is the radar's score: discussion, source authority and the lab's own posts, fading with age.

  1. DiscussionThinking Machines: Inkling Smalltwitter.com · 2026-07-30 · 4 HN pts

Where to run it

Live hosted endpoints, cheapest first, USD per million tokens. Ties break on output price, then provider name.

#Cheapest 2In /1MOut /1MContextQuantSeen
1 DeepInfra $0.45 $1.20 512K fp8 Oct 25
2 Thinking Machines · free tier $0 $0 1M nvfp4 Oct 24

All 2 providers →

Where it's used

Hugging Face does not say who downloads a model or with what. These are the proxies: API providers listing it, quantized builds, fine-tunes/adapters/merges, and Spaces that load it.

Served by 2 API providers; heavily run locally (56 quantized builds); a big base for builders (56 fine-tunes, adapters and merges).

  • As a service2DeepInfra, Thinking Machines
  • On own hardware56quantized builds
  • Built on567 fine-tunes · 49 adapters · 0 merges
  • In apps23Hugging Face Spaces

Top Spaces running it: HuggingFaceM4/hf-realtime-voice-inkling ♥ 14 · ArtShumov/Whyx-PROmpTea ♥ 3 · suvadityamuk/blender-agent ♥ 1 · localellm/WebGPU-inkling-small-gguf ♥ 1

Will it fit?

Weights only, an estimate: leave room for the KV cache. Method

Fits two 80 GB GPUs at INT4; too big for one

About 160 GB at INT4, 319 GB at FP8, 638 GB at BF16: weights only, so leave room for the KV cache. Every expert must be resident, so total parameters count.

16 GB card64 GB Mac2×80 GB8×80 GBINT4: about 160 GB of weights (estimate; KV cache and activations extra)INT4160 GBFP8: about 319 GB of weights (estimate; KV cache and activations extra)FP8319 GBBF16: about 638 GB of weights (estimate; KV cache and activations extra)BF16638 GB
16 GB card64 GB Mac2×80 GB8×80 GBINT4: about 160 GB of weights (estimate; KV cache and activations extra)INT4160 GBFP8: about 319 GB of weights (estimate; KV cache and activations extra)FP8319 GBBF16: about 638 GB of weights (estimate; KV cache and activations extra)BF16638 GB

Estimate params × bytes per param × 1.2 · how we estimate

The story so far

Weights, every dated article, and repo commits on one rail.

Most coverage came in the first two weeks

Teal is the weights (2026-07-27), amber triangles are 3 dated articles (each links to it), grey ticks are repo commits.

AugSepOct2026-07-31: 1 repo commit2026-07-30: 9 repo commits2026-07-27: 2 repo commitsInkling-Small · 2026-07-30Inkling Small is the highest-scoring open-weight model evaluated by ARC Prize · 2026-07-30Thinking Machines: Inkling Small · 2026-07-30Weights 2026-07-27 · 2026-07-27Weights 2026-07-27now · 2 providers
AugSepOct2026-07-31: 1 repo commit2026-07-30: 9 repo commits2026-07-27: 2 repo commitsInkling-Small · 2026-07-30Inkling Small is the highest-scoring open-weight model evaluated by ARC Prize · 2026-07-30Thinking Machines: Inkling Small · 2026-07-30Weights 2026-07-27 · 2026-07-27Weights 2026-07-27now · 2 providers

What people built on it

112 derivatives on the Hub. Which build to run, by runtime: the most downloaded in each class.

112 derivatives on the Hub, most of them quantized

From the Hugging Face model tree; quantized builds in teal, fine-tunes, merges and adapters in greys.

56 quantized (Hugging Face model tree)7 fine-tunes (Hugging Face model tree)49 adapters (Hugging Face model tree)56 quantized7 fine-tunes49 adapters
56 quantized (Hugging Face model tree)7 fine-tunes (Hugging Face model tree)49 adapters (Hugging Face model tree)56 quantized7 fine-tunes49 adapters
GGUF unsloth/Inkling-Small-GGUF for llama.cpp, Ollama, LM Studio 2.9M downloads MLX mlx-community/Inkling-Small-mxfp4 for Apple silicon 2.1k downloads NVFP4 thinkingmachines/Inkling-Small-NVFP4 for vLLM, SGLang 1.2M downloads

All variants →

Reference

Everything else, folded: the full tables, every source.

Benchmarks

8 rows

Benchmarks 8 rows

measured third-party leaderboard rowauthor-reported number from the model card or eval files (8)No confidence intervals unless the source gave them; settings differ between rows, so compare within a benchmark, not across. Date = the benchmark's own date when it gives one, else when we read it.
BenchmarkValueSetting / notesReported byDate
author-reported SWE-bench Verified · Write code with agents80.2% percent—model author2026-10-07
author-reported MathArena/aime_202695.5—Model Card2026-10-24
author-reported diamond (Idavidrein/gpqa)89.5—Model Card2026-10-24
author-reported hle (cais/hle)31.6text onlyModel Card2026-10-24
author-reported MathArena/hmmt_feb_202690.2—Model Card2026-10-24
author-reported mmmu_pro_standard_10_options (MMMU/MMMU_Pro)74—Model Card2026-10-24
author-reported SWE_Bench_Pro (ScaleAI/SWE-bench_Pro)55.9—Model Card2026-10-24
author-reported swe_bench_%_resolved (SWE-bench/SWE-bench_Verified)80.2—Model Card2026-10-24

What the card says

2 cited sentences

Summary generated, cited

Inkling-Small is a 266B parameter multimodal autoregressive transformer designed for text, image, and audio inputs.[1][2] The model uses a sparse Mixture-of-Experts architecture with a 42-layer decoder-only backbone.[3][4]

  1. parameters human
  2. “Multimodal autoregressive transformer”
  3. is mixture of experts
  4. “A 42-layer decoder-only transformer with a sparse Mixture…”

Generated 2026-10-07 15:09 UTC by parasail-gemma-4-26b-a4b-it; each numeral points at the fact it rests on. Not a verdict.

Use-case evidence

6 use cases

What it is good at 6 use cases

Labels are evidence, not verdicts: measured a third-party leaderboard row for this model; author-reported a number the model author published in the card or eval files; structural a deterministic signal from the name, config or chat template, or a decision-model estimate. Each chip links to its source. All use cases.

Facts

21 fields

Listed in LLMs · MoE

Facts

Pipeline tagimage-text-to-text
Librarytransformers
ArchitecturesInklingForConditionalGeneration
Model typeinkling_mm_model
Dtypebfloat16
Params by dtypeBF16: 265,956,428,810 · F32: 10,280
Size bandXXL
MoEyes
Experts256
Experts per token6
Context length1,048,576 (provider listing)
LicenseApache 2.0 (apache-2.0)
Hub snapshot2026-10-25

Servability

vLLM · SGLang

Servability

vLLM: supported · InklingForConditionalGeneration · 7050ef881ff0835f6890a47db1859a24c65481d6SGLang: supported · InklingForConditionalGeneration · 773cb938fe9484a8f7e7e46fb1b1fe96412c59ed

vLLM registry (7050ef881ff0835f6890a47db1859a24c65481d6) · SGLang registry (773cb938fe9484a8f7e7e46fb1b1fe96412c59ed) · Matched by architecture class name against each project's registry on main; 'supported' means the architecture is registered, not that this checkpoint was tested.

Hardware fit

3 estimates

Hardware fit ESTIMATE

Weights-only memory for Inkling-Small: params × bytes/param × 1.2 overhead. KV cache and activations are extra. How this is computed.

BF16 (2 bytes/param)638 GB · Multi-GPU node (<640 GB)
FP8 (1 byte/param)319 GB · Multi-GPU node (<640 GB)
INT4 (0.5 bytes/param)160 GB · Multi-GPU node (<640 GB)

Weights only (params × bytes/param × 1.2 overhead); KV cache and activations extra. MoE: all experts must be resident, so total parameters count.

All variants

40 variants

Lineage

Original model Family: thinkingmachines/Inkling-Small 112 derivatives on the Hub

quantized: 56finetune: 7adapter: 49

Top variants (40 shown, ranked by downloads)

Relation
Quant
Size
40 rows
Variant Relation Quant Params Band License Author Likes Downloads
unsloth/Inkling-Small-GGUF quantized GGUF — — apache-2.0 unsloth 98 2.9M
thinkingmachines/Inkling-Small-NVFP4 quantized NVFP4 156B XXL apache-2.0 thinkingmachines 88 1.2M
miweru/Inkling-Small-REAP-137B-A12B-de-GGUF quantized GGUF — — apache-2.0 miweru 4 7.7k
Baekpica/Inkling-Small-Mixed-Quant-GGUF quantized -Quant — — apache-2.0 Baekpica 2 7.1k
lyssquant/Inkling-Small-GGUF quantized GGUF — — apache-2.0 lyssquant 0 5.1k
RedHatAI/Inkling-Small-FP8-dynamic quantized -FP8 266B XXL apache-2.0 RedHatAI 2 3.9k
Baekpica/Inkling-Small-GGUF quantized GGUF — — apache-2.0 Baekpica 0 3.2k
cyankiwi/Inkling-Small-AWQ-INT4 quantized AWQ 266B XXL apache-2.0 cyankiwi 4 2.9k
mlx-community/Inkling-Small-mxfp4 quantized mxfp4 264B XXL — mlx-community 4 2.1k
DevQuasar/thinkingmachines.Inkling-Small-GGUF quantized GGUF — — — DevQuasar 0 2.2k
xautonomics/Inkling-Small-UD-Q4_K_M.LAYR.GGUF quantized GGUF — — apache-2.0 xautonomics 0 2.1k
mlx-community/Inkling-Small-mlx-2bit finetune mlx 264B XXL apache-2.0 mlx-community 3 1.4k
unsloth/Inkling-Small finetune — 266B XXL apache-2.0 unsloth 9 583
pipenetwork/Inkling-Small-MLX-4bit quantized MLX 264B XXL apache-2.0 pipenetwork 2 920
mlx-community/Inkling-Small-mlx-4bit finetune mlx 264B XXL apache-2.0 mlx-community 1 920
pipenetwork/Inkling-Small-MLX-REAP25-4bit quantized MLX 199B XXL apache-2.0 pipenetwork 2 765
Sawfwair/Inkling-Small-MLX-Mixed-2bit quantized MLX 264B XXL apache-2.0 Sawfwair 2 724
0xSero/Inkling-Small-EXL3-2.0bpw quantized EXL3 40.7B XL apache-2.0 0xSero 0 696
mlx-community/Inkling-Small-mlx-3bit finetune mlx 264B XXL apache-2.0 mlx-community 1 619
inferencerlabs/Inkling-Small-MLX-Q9 quantized MLX — — — inferencerlabs 1 608
pipenetwork/Inkling-Small-MLX-3bit quantized MLX 264B XXL apache-2.0 pipenetwork 1 508
AxionML/Inkling-Small-NVFP4 quantized NVFP4 156B XXL apache-2.0 AxionML 0 441
cebopp1/inkling-tiny-preview adapter — 14.0B L apache-2.0 cebopp1 0 428
ToPo-ToPo/Inkling-Small-mlx-2bit quantized mlx 264B XXL apache-2.0 ToPo-ToPo 0 419
EmbeddedLLM/Inkling-Small-MXFP4 quantized MXFP4 148B XXL apache-2.0 EmbeddedLLM 0 406
NextTokenAI/NextSearch-1-M finetune — 266B XXL apache-2.0 NextTokenAI 0 403
ToPo-ToPo/Inkling-Small-mlx-4bit quantized mlx 264B XXL apache-2.0 ToPo-ToPo 0 378
lightseekorg/Inkling-Small-MXFP4 quantized MXFP4 148B XXL apache-2.0 lightseekorg 1 304
cebopp1/inkling-tiny-reasoning-v1 adapter GGUF — — other cebopp1 1 303
sabrewing-engine/Inkling-Small-colibri-int4 quantized int4 — — apache-2.0 sabrewing-engine 1 195
olka-fi/Inkling-Small-MXFP4 quantized MXFP4 282B XXL apache-2.0 olka-fi 0 211
0xSero/Inkling-Small-EXL3-2.5bpw quantized EXL3 48.7B XL apache-2.0 0xSero 1 154
INCModel2/Inkling-Small-MXFP4-CT-AutoRound quantized MXFP4 266B XXL apache-2.0 INCModel2 1 152
0xSero/Inkling-Small-EXL3-3.0bpw quantized EXL3 56.8B XL apache-2.0 0xSero 0 122
0xSero/Inkling-Small-EXL3-4.5bpw quantized EXL3 80.9B XL apache-2.0 0xSero 0 122
0xSero/Inkling-Small-EXL3-4.0bpw quantized EXL3 72.9B XL apache-2.0 0xSero 0 121
0xSero/Inkling-Small-EXL3-5.0bpw quantized EXL3 89.0B XL apache-2.0 0xSero 0 115
0xSero/Inkling-Small-EXL3-3.5bpw quantized EXL3 64.8B XL apache-2.0 0xSero 0 109
fribourgquant/Inkling-Small-NVFP4 quantized NVFP4 156B XXL apache-2.0 fribourgquant 0 97
shaffhausen/Inkling-Small-NVFP4 quantized NVFP4 156B XXL apache-2.0 shaffhausen 0 93

Repository changelog

12 commits
changed in the last 7 daysOther: 2026-10-24all changes

Repository changelog 12 commits

DateCommitSHA
2026-07-31 Add community evaluation results for AIME_2026, GPQA, HLE, HMMT_FEB_2026, MMMU_PRO, SWE-BENCH_PRO, SWE-BENCH_VERIFIED (#2) 8cc5877
2026-07-30 Update README.md b2d4f22
2026-07-30 Update README.md 4024fac
2026-07-30 Update README.md 173638a
2026-07-30 Update README.md d22e7eb
2026-07-30 Update README.md 28da9a6
2026-07-30 Update README.md faeaccf
2026-07-30 Update README.md ccd03dc
2026-07-30 Update README.md 191421c
2026-07-30 Update README.md 9881ccf
2026-07-27 Upload Inkling-Small model 21152b5
2026-07-27 initial commit 8de0d24

All providers

2 providers · 2 endpoints

Where to run 2 providers · 2 endpoints

Hosted endpoints observed for Inkling-Small, sorted by input price (USD per million tokens). Each row names its observer and date; prices change daily, so check the source before quoting. Cheapest observed input price: $0.00/M.

ProviderModel refContextQuant$/M in$/M outtok/sTTFTToolsStatusObserved
Thinking Machines
openrouter
 thinkingmachines/inkling-small 1M nvfp4 $0.00 $0.00 — — yes live 2026-10-24
DeepInfra
hf, deepinfra, openrouter
 thinkingmachines/Inkling-Small 512K fp8 $0.45 $1.20 142.7 513 ms yes live 2026-10-25

Articles and coverage

3 pieces

Ranked by source signals (HN, Lobsters, authority domains, official posts) and recency. Summaries are generated from a stored copy of the article text; every sentence cites a span of that text, and nothing is shown that failed the citation check.

Analysis1

Inkling-Small
thinkingmachines.ai2026-07-30HN 33 pts · 2 comments2,416 words
Thinking Machines releases Inkling-Small, an efficient open-weights Mixture-of-Experts model with 276B total parameters and 12B active trained on NVIDIA GB300 NVL72 systems. The model features a context window of up to 1M tokens and achieves comparable performance to Inkling at a quarter of its size across agentic tool use (Terminal-Bench 2.1), reasoning (HLE text-only), and instruction following (IFBench) benchmarks. It is available for fine-tuning on Tinker and for text, image, and audio chat on Tinker Playground, with variable thinking effort to balance cost and performance.written, reviewed, cited
  • Inkling-Small surpasses Inkling on reasoning and agentic coding benchmarks despite Inkling maintaining an advantage on knowledge coverage and factuality.
  • The model uses an encoder-free architecture where images are divided into 40×40-pixel patches and transformed using a four-layer hMLP.
  • ForecastBench and Prophet Arena results were obtained during testing between July 19 and July 28, 2026.

Community2

Inkling Small is the highest-scoring open-weight model evaluated by ARC Prize
twitter.com2026-07-30HN 6 pts · 0 comments
Thinking Machines: Inkling Small
twitter.com2026-07-30HN 4 pts · 0 comments

Similar models

3 models

Similar models

  • GLM-5.3-Flash · 321B
    same size band (XXL), also MoE, same task (image-text-to-text), more providers, permissive license
  • Qwen3-VL-235B-A22B-Instruct · 236B
    same size band (XXL), also MoE, same task (image-text-to-text), more providers, permissive license
  • Qwen3-VL-235B-A22B-Thinking · 236B
    same size band (XXL), also MoE, same task (image-text-to-text), more providers, permissive license

Sources

4 sources

Sources

Hub snapshot: 2026-10-25 · HF API: 2026-10-24 21:39 UTC · Providers: 2026-10-24 21:44 UTC · Articles: 2026-10-24 21:46 UTC · Summary: 2026-10-07 15:09 UTC · Last verified 2026-10-25

Follow Inkling-Small

Watch this model

Email alerts are not open yet; the feeds above are the same data. See feeds for the item schema.