Thinking Machines and third parties publish details on 4-bit training approach and efficiency tradeoffs
Thinking Machines Inkling 2 Small
Latest: Thinking Machines' 594B Inkling 2 Small Available Free on OpenRouter
A 594B open-weight model under Apache 2.0 is accessible to a wide range of developers and researchers without licensing restrictions. Benchmark results show meaningful gaps across task types, so the model's practical fit depends heavily on the application.
When stories like this move, I write about them in an email each morning.
One free email each morning, around 8 AM US Eastern.
Check your email.
I have sent you a confirmation link. One click and you are in.
Not in your inbox? Look in spam or promotions. If it's there, mark it "Not spam" so future editions land in your inbox.
The full picture
Inkling 2 Small is available at no cost on OpenRouter for a limited time, compatible with agentic coding harnesses including Claude Code, Codex, and Hermes Agent, and pairs vision with agentic coding workflows for building websites, games, and 3D designs. Thinking Machines released the model on November 5 with weights on HuggingFace under Apache 2.0. The model has 594 billion parameters, was pretrained on 8.4 trillion tokens using NVIDIA GB300 NVL72 racks in NVFP4 format, and requires at least 500GB GPU memory in NVFP4 or 800GB in MXFP8. Training used a native 4-bit forward pass from pretraining through final RL rollout, cutting pretraining time by 8% and RL time by 20% at roughly 2% efficiency loss, and because numerics remained consistent throughout training the released checkpoint required no separate calibration step. In an internal benchmark, the model autonomously trained Qwen3.5-4B on time series prediction tasks without human intervention, completing the work in 4.5 hours. Benchmark results are uneven: the model outperforms Mistral Large 4 and Muse Spark Open on DeepSWE, JobBench, Finance Agent, and LongBench v2, and its internal misbehavior score dropped from 75.2 to 14.9 versus the prior Inkling Small, though GPT-6.1 Sol scored 0.8 on that same metric and Muse Spark Open scores higher on AA-Omniscience. MindStudio testing found it scored 13.1% on terminal-bench style coding tasks versus DeepSeek V4.1 Flash's 33.8%, though it outperforms that model on vision, knowledge work, and finance evaluations. Vals AI ranked it 19th of 34 overall at 56.08%, with strong scores on SAGE (88.57%, rank 8 of 29), Vibe Code Bench (81.55%), Legal Index (73.89%), and Finance Agent v2.1 (67.54%, rank 4 of 37), but near-bottom scores on SRE Bench (1.91%) and ProgramBench (0.50%). The model supports a 1M token context window and is priced at $0.30 per million input tokens and $1.00 per million output tokens.
How it developed
Thinking Machines released Inkling 2 Small on November 5 with weights on HuggingFace under Apache 2.0, then confirmed it free on OpenRouter and compatible with Claude Code, Codex, and Hermes Agent.
Benchmark results placed it 19th of 34 in Vals AI's index, with strong finance and legal scores but near-zero on SRE Bench and ProgramBench. VentureBeat reported it outperformed Mistral Large 4 on enterprise tasks including DeepSWE and Finance Agent.
MindStudio publishes hands-on testing showing strong vision and knowledge-work performance but weaker coding benchmark scores
VentureBeat reports benchmark results showing outperformance on enterprise tasks and a sharp drop in misbehavior score
Thinking Machines announces Inkling 2 Small, a 594B model pretrained on 8.4T tokens using NVIDIA GB300 NVL72 racks in NVFP4 format
Sources
1 more source
- Alibaba's Qwen3 Max open-source release
- DeepSeek V4.1-Flash-1017 open-weight release
- DeepSeek V4-Pro open-weight release
- Meta's Muse Ember small model launch
BackgroundFrontier models, open models, and how they are compared
Get the daily email
When stories like this move, I write about them in an email each morning.
One free email each morning, around 8 AM US Eastern.
Check your email.
I have sent you a confirmation link. One click and you are in.
Not in your inbox? Look in spam or promotions. If it's there, mark it "Not spam" so future editions land in your inbox.