Inkling: Thinking Machines Lab's Open-Weights MoE Model — Benchmarks, Architecture, and What It Actually Competes On
Thinking Machines Lab (the AI company founded by former OpenAI CTO Mira Murati) dropped their first Inkling open-weights model yesterday. Inkling is a 975-billion parameter Mixture-of-Experts transformer with 41 billion active parameters per token, native multimodal support (text, images, audio, video), up to 1M tokens of context, and controllable reasoning effort. The weights are on Hugging Face.
The headline number that got everyone's attention: 935 points and 229 comments on Hacker News within the first day. That's the kind of reception you get when an American company finally ships a genuinely competitive open-weights model after the Chinese labs (DeepSeek, Z.ai/GLM) have been dominating the narrative.
But after reading the full announcement, the benchmark table, and the HN thread, the honest take is more nuanced than "Inkling beats everyone." Let me walk through what's actually impressive, where it falls short, and whether this matters for how you build.
Inkling Open-Weights Model Architecture: MoE with Interesting Choices
Inkling follows the DeepSeek-V3 playbook for its MoE layer design but makes several distinct choices:
- 256 routed experts + 2 shared experts, 6 routed experts active per token
- Sigmoid-based router with auxiliary-loss-free load-balancing bias (no balancing loss that degrades performance)
- Relative positional embeddings instead of the now-ubiquitous RoPE — they claim it performs better and extrapolates better to longer sequences
- 5:1 sliding-window to global attention layers with 8 KV heads
- Short convolutions after KV projections and on residual branch outputs before rejoining the main stream
The relative-positional-embedding choice is worth noting. Almost every major model today uses RoPE (introduced in the GPT-NeoX-20B / EleutherAI era). Thinking Machines went with the older Shaw-style relative embeddings and short convolutions instead. They trained on 45 trillion tokens across modalities and the architecture held up fine. That's a useful data point — RoPE is excellent but it's not the only viable approach for long-context models.
Training used a hybrid Muon + Adam optimization strategy with weight decay coupled to the square of the learning rate — a technique from their modular manifolds research. The hardware was NVIDIA GB300 NVL72 systems, so this wasn't a cheap training run.
Inkling-Small (276B total, 12B active) was also previewed and performs shockingly close to its larger sibling on reasoning and agentic tasks while being a fraction of the cost.
The Benchmark Reality: Honest but Not Frontier
The official announcement is refreshingly candid: "Inkling is not the strongest overall model available today, open or closed."
Here's Inkling compared to GPT 5.6 Sol, Claude Fable 5, GLM 5.2, and Nemotron 3 Ultra on the key axes:
| Benchmark | Inkling | GLM 5.2 | GPT 5.6 Sol | Claude Fable 5 |
|---|---|---|---|---|
| AIME 2026 (Math Reasoning) | 97.1% | 99.2% | 99.9% | 99.9% |
| HLE (no tools) (Hard Reasoning) | 29.7% | 40.1% | 47.2% | 53.3% |
| SWE-bench Pro Public (Coding) | 54.3% | 62.1% | 64.6% | 80.0% |
| Terminal Bench 2.1 (Agentic Coding) | 63.8% | 82.7% | 89.5% | 84.6% |
| MCP Atlas (Agentic General) | 74.1% | 77.8% | 81.8% | 83.3% |
| MMMU Pro (Vision) | 73.5% | — | 83.0% | 84.2% |
| FORTRESS Adversarial (Safety) | 78.0% | 71.3% | 82.4% | 96.0% |
The trend is clear: Inkling is competitive on math reasoning (97.1% on AIME), decent on agentic/coding tasks (in the ballpark of Nemotron and GLM 5.2 on SWE-bench), but noticeably behind the frontier closed models on hard reasoning and agentic benchmarks.
On agentic coding specifically — Terminal Bench 2.1 at 63.8% vs GLM 5.2's 82.7% is a material gap. The HN discussion noted that GLM 5.2 went through extensive post-training iteration since its original release, while Inkling is a first release with room to grow.
Where Inkling genuinely wins: multimodal breadth and fine-tunability. It's one of the few open-weights models that natively handles text, images, audio, and video in one architecture with controllable thinking effort. Neither GLM 5.2 nor Nemotron have native audio. Inkling scores 56.6% on Audio MC and 91.4% on VoiceBench.
The Self-Fine-Tuning Demo Is the Most Interesting Part
The announcement includes an interactive terminal demo where Inkling fine-tunes itself using Tinker. The model wrote its own fine-tuning job (a lipogram task — never using the letter "e"), generated the training data, ran the Tinker post-training API, and self-updated its weights in OpenCode. The whole loop took about 27 minutes.
That's the pitch: open weights + Tinker fine-tuning platform. The business model is "give away the weights, charge for the SFT/RL infrastructure." HN commenters debated whether this works — you could self-host and fine-tune independently — but the convenience argument is real for teams that don't want to manage GPU infrastructure.
The controllable thinking effort is also genuinely useful. Inkling has 3 modes (non-thinking, high, max) that let you trade latency for reasoning depth. At effort=0.99, you get full reasoning; at non-thinking, maximum throughput. The RL training data shows reasoning performance improved log-linearly over 30M+ rollouts, and the chain of thought became more compressed and efficient over time — an emergent property of the reward function.
Competition Landscape: The US Open-Weights Gap
This release lands in a very different market from when Thinking Machines was founded:
- DeepSeek-V4 and GLM 5.2 have set a high bar for open-weights models from Chinese labs
- GPT 5.6 Sol / Terra / Luna and Claude Fable 5 dominate the closed-source frontier
- Meta reportedly pivoted away from Llama 4 after disappointing results
- AllenAI has yet to produce a model at this scale
Inkling slots into the "good open-weights base for customization" tier. It's not the best general-purpose model, but it's multimodal, open-weights, and available on a fine-tuning platform. That combination doesn't have many competitors in the American ecosystem right now.
One fascinating HN thread: "There is no DeepSeek of the US" because Chinese labs enjoy state support for open-weight releases as strategic assets, while American companies face shareholder pressure to monetize. Thinking Machines, as a public company (they IPO'd recently), is trying to thread that needle with a novel business model.
When to Use Inkling
Use it if:
- You need an open-weights model with native multimodal support (especially audio)
- You want to fine-tune on proprietary data and Tinker makes that easy
- You want controllable thinking effort for your agentic workflows
- You need self-hostable weights (NVFP4 checkpoint available for Blackwell)
Skip it if:
- You need frontier-level coding agent performance (Claude Fable 5 dominates here)
- You're optimizing purely for cost on agentic coding (GLM 5.2 API is cheaper and stronger on Terminal Bench)
- Your workflow is text-only and you don't need audio
Inkling-Small at 12B active parameters with near-Inkling reasoning performance is the more interesting long-term play, but it's still in preview.
Hard Limitations
- Factuality is weak: SimpleQA Verified at 43.9% is well below GPT-5.6 Sol (71.6%) and Claude Fable 5 (68.3%). Verify outputs carefully.
- Agentic coding gap: 54.3% SWE-bench Pro and 63.8% Terminal Bench 2.1 trail the leaders significantly. Not a Cursor/Claude Code replacement yet.
- Self-hosting is expensive: 975B parameters means 1.51TB in BF16. Even 4-bit quantization needs ~375GB VRAM (5+ H100s). Cloud/enterprise only.
- First release risk: Thinking Machines acknowledges this is their first major training effort. Inkling will likely improve, but the benchmarks are what they are today.
Top comments (0)