Atlas

Models

Thinking Machines LabOpen weights

Inkling 2 Mini

155.6
Capability · 95% interval 154.3–157.1
Rank 90 of 901 (interval 85–100)
Relative cost 0.18× · $0.011 per short-answer task · $0.19 per agentic task

Inkling 2 Mini is Thinking Machines Lab's open-weight multimodal looped Mixture-of-Experts model. Its model card reports 26.4 billion total and 3.4 billion active parameters, and the released safetensors hold 26,388,641,280 tensor parameters, not counting block scales or the separate multi-token-prediction head. The 24-layer decoder is split into four blocks of six layers, and the two middle blocks run three times with shared weights, so each token takes 48 layer applications and 5.4 billion weight activations. Each token is routed to 8 of 128 experts, left after two-thirds of the original 384 were pruned in midtraining and the model was healed on 1 trillion distilled tokens, and attention mixes 1,024-token sliding-window and global layers with grouped-query attention. It accepts text and image, supports variable reasoning effort from 0.1 to 0.99, though the card says the minimum doesn't fully disable reasoning, and a context window of up to 1M tokens. Twenty layers were trained and ship natively in NVFP4 and the last four in MXFP8, in a single checkpoint of about 20GB. Thinking Machines Lab released it on Hugging Face under Apache 2.0 and on Tinker on November 20, 2026. Its launch post and model card don't state the effort, temperature or token limits their evaluations ran at.

Scored on 217 benchmarks · 216 total scores · 188 independently reported

Score providers

Organizations reporting headline scores for 217 benchmarks

Inkling 2 Mini benchmark scores

In fitin the capability model (52 of 217)
Professional · 117 benchmarks · first place on 1
AA-AnalystAgent (pass@5)not reported
1 / 54———
AA-AnalystAgent (pass@1)not reported
2 / 54———
HLAB-Vals-1.1 - White Collar Defense Investigations0.99
2 / 42———
HLAB-Vals-1.1 - Employment Labor0.99
3 / 42———
JobBench - release-reported aggregation unspecifiednot reported
6 / 15——2026-11-20
Software · 46 benchmarks
PMPP-Hard (Thinking Machines 10-hour AUC)not reported
2 / 2——2026-11-20
CyberBench v1.1 Patch0.99
4 / 39———
SWE-bench Multimodal (Vals v1): Scratch GUI0.99
7 / 21———
CyberBench v1.1 OverallIn fit0.99
9 / 38~11 mo2028–2031—
SWE-bench Multimodal (Vals v1): Grommet0.99
12 / 21———
Knowledge & reasoning · 14 benchmarks
LongBench v2In fitnot reported
6 / 55~3 mo~Aug 20282026-11-20
OpenAI MRCR v2 8-needle - 256K-512KIn fitnot reported
8 / 9~7 moApr 20262026-11-20
GraphWalks BFS - 256KIn fitnot reported
10 / 12~4 moJun 2026*2026-11-20
AA-Omniscience v1.1 Hallucination Ratenot reported
41 / 85———
Humanity's Last Exam Diamond (Reasoning Partition)not reported
45 / 109———
Science · 12 benchmarks
Terminal-Bench Science — Engineering Sciences (Vals)0.99
21 / 46———
Vals Science Index0.99
25 / 26———
Terminal-Bench Science — Earth Sciences (Vals)0.99
27 / 46———
BioMysteryBench (Vals Terminus 2)In fit0.99
32 / 33~2 mo~Jun 2027—
BioMysteryBench — Human-solvable (Vals)0.99
32 / 33———
Perception & media · 10 benchmarks
ScreenSpot-Pro (with tools)not reported
3 / 11——2026-11-20
MuirBenchIn fitnot reported
5 / 29~8 mo~May 20272026-11-20
BabyVision with Pythonnot reported
6 / 22——2026-11-20
PerceptionBenchIn fitnot reported
6 / 20~6 mo2027–20302026-11-20
ScreenSpot-ProIn fitnot reported
7 / 55~1.4 yearsFeb 20262026-11-20
Tool & computer use · 7 benchmarks
MCP Atlas April 2026 (public 500 tasks)not reported
3 / 11——2026-11-20
DeepSearchQA (F1 score)In fitnot reported
17 / 35~6 moFeb 20262026-11-20
SkillsBench without SkillsIn fit0.99
18 / 47~2 mo~Feb 2028—
EnterpriseOps-Gym-AAIn fitnot reported
25 / 53~1.7 yearsNo trend—
BrowseCompIn fitnot reported
31 / 80~5 moSep 20262026-11-20
Math · 4 benchmarks
Apex (MathArena)In fitnot reported
12 / 53~2 moJul 2026*2026-11-20
MysteryMechanismIn fit0.99
22 / 33~2 mo~Mar 2028—
Terminal-Bench Science — Mathematical Sciences (Vals)0.99
30 / 46———
ProofBench v1.1In fit0.99
42 / 52~4 moFeb 2026—
Composite indexes · 3 benchmarks
Vals Index v2.20.99
33 / 37———
Artificial Analysis Intelligence Index v5.0.1not reported
43 / 76———
Capability0.99
155.6 / 171Capability
90 / 901——2026-11-20
Health & medicine · 2 benchmarks
MedScribeIn fit0.99
24 / 98~8 moSep 2026—
MedCodeIn fit0.99
94 / 98~7 mo2029–2031+—
Safety & alignment · 2 benchmarks · first place on 1
StrongREJECT (no jailbreak) - release-reported scorenot reported
1 / 2——2026-11-20
HoneyBench v0.2 (hack rate)not reported
6 / 19——2026-11-20

Related models

Loading Atlas data…