Inkling: Thinking Machines Lab Open-Weights MoE for Customization (July 2026)
Jul 15, 2026: Inkling — 975B MoE, 41B active, 1M context, multimodal, Tinker fine-tuning. SWE-bench 77.6%, effort dial 0.2–0.99. Not #1 overall — best open base for customization. explainx.ai maps benchmarks, cost, and how to run it.
On July 15, 2026, Thinking Machines Lab released Inkling — its first model trained from scratch with full open weights. Five days after the lab's "Future Worth Building Is Human" manifesto, Inkling is the bet that customizable multimodal foundations matter more than winning every public leaderboard.
By July 16, the release was trending on X with roughly 9,200 posts in Grok's topic feed — developers asking the same three questions: Is this the new #1 open model? Can I actually fine-tune it? And why does the team say it is not the strongest overall?David Siegel's Fortune piece (Jul 3) names exactly that tension: open weights you can run and customize vs frontier closure and missing training-pipeline transparency.
No — Thinking Machines says Inkling is the best open-weights base for customization, not peak scores everywhere. Strong on breadth and multimodality; trails Kimi K2.6 and GLM 5.2 on several agentic coding rows.
HLE text 29.7% / with tools 46% · SWE-bench Verified 77.6% · Terminal Bench 2.1 63.8% · Design Arena Agentic Web Dev ~1253 (open-weights tier) · FORTRESS adversarial 78% · StrongREJECT 98.6%.
What is the effort dial?
0.2–0.99 — lower effort = fewer reasoning tokens. Matches Nemotron 3 Ultra on Terminal Bench 2.1 at ~⅓ tokens per Thinking Machines' sweep charts.
How do I get weights?
Hugging Face full checkpoint + NVFP4 variant. APIs via Together, Fireworks, Modal, Databricks, Baseten. Inference: SGLang, vLLM, llama.cpp.
How do I customize?
Tinker — 64K/256K context, 50% discount for a limited launch window. Cookbook recipes + Inkling Playground for chat before you commit GPU hours.
What harness did the demo use?
OpenCode — Inkling fine-tuned itself into a lipogram (no letter "e") via Tinker inside the OpenCode agent loop. ~27 minutes end-to-end.
Is there a smaller model?
Inkling-Small preview: 276B total, 12B active. Close on reasoning; weaker on Terminal Bench and factuality. Full weights pending testing.
What Inkling is — and what it is not
Thinking Machines' positioning is unusually honest for a launch week:
"Inkling is not the strongest overall model available today, open or closed. Instead, a combination of qualities makes it a good open-weights base for customization: multimodal capabilities, efficient thinking, and availability on Tinker for fine-tuning."
That framing matters if you are shopping for a Fable 5 replacement or a Kimi K2.7 coding default. Inkling is a broad foundation — agentic coding, reasoning, instruction following, vision, audio, safety, and forecasting calibration — trained so organizations can specialize it, not so benchmark Twitter can crown a single winner.
The lab's July 10 manifesto argued that alignment and capability should live in weights you own. Inkling ships that argument as a 975B-parameter artifact: pretrained from scratch on GB300 NVL72 clusters, post-trained with 30M+ RL rollouts, bootstrapped partly from Kimi K2.5 synthetic data for initial SFT — then scaled with Muon+Adam optimization and large-scale asynchronous RL.
If you need a mental model: Inkling is closer to "open Nemotron-class generalist you can reshape" than "Kimi K2.7 coding specialist out of the box." For the coding-specialist lane, see explainx.ai's Kimi K2.7 guide and K3 beta leaks.
Architecture and training — why the recipe is different
MoE and long context
Inkling follows the DeepSeek-V3 MoE pattern: each layer has 256 routed experts + 2 shared experts, with 6 routed experts active per token. Routing uses a sigmoid router with auxiliary-loss-free load balancing.
For attention, Thinking Machines interleaves sliding-window and global layers at 5:1 with 8 KV heads — a cost control for 1M-token contexts. Notably, they use relative positional embeddings instead of RoPE, citing better extrapolation on long sequences. Short convolutions sit after K/V projections and on residual branches.
Multimodal: encoder-free audio and vision
Inkling is natively multimodal — not a text model with bolt-on encoders:
Chart/diagram reasoning; optional Python zoom/crop tool at inference
On audio benchmarks at effort 0.99, Inkling reports VoiceBench 91.4%, MMAU 77.2%, Audio MC 56.6% — competitive with specialist omni open models like Qwen3-Omni and Nemotron-3-Nano-Omni, though closed Gemini 3.1 Pro still leads several vision rows.
Multimodal training aligns with the lab's interaction models preview — voice-and-vision collaboration in real time, not batch upload pipelines.
RL at scale and emergent efficiency
The training chart Thinking Machines published is instructive: aggregate reasoning reward (AIME, HLE, GPQA, etc.) improves log-linearly from SFT init (0.264) through 30M+ rollouts to the released checkpoint (0.356).
Two emergent behaviors surfaced during RL:
Controllable thinking effort — the lab varied system messages and per-token costs during rollouts so the model learned to spend more or fewer tokens by task.
Compressed chain-of-thought — reasoning traces dropped grammatical overhead over training ("We need to understand" → "We need determine") without hurting final answers — similar to what Cognition reported training SWE-1.7.
That efficiency story is Inkling's real pitch for agent harnesses where token burn dominates COGS. If you are comparing harness overhead itself, see Claude Code vs OpenCode token study.
Benchmark map — effort 0.99, temperature 1.0
All numbers below come from Thinking Machines' July 15 tables unless noted. Coding evals use a 256K max-token trajectory limit. For how to read these suites, start with explainx.ai's AI benchmarks complete guide — especially the SWE-bench contamination and harness-variance sections.
Reasoning and agentic coding
Benchmark
Inkling
Nemotron 3 Ultra
Kimi K2.5
Kimi K2.6
GLM 5.2
Claude Fable 5 (max)
GPT 5.6 Sol (max)
HLE (text only)
29.7%
26.6%
29.4%
35.9%
40.1%
44.7%
47.2%
HLE (with tools)
46.0%
37.4%
50.2%
54.0%
54.7%
51.4%
55.0%
SWE-bench Verified
77.6%
70.7%
76.8%
80.2%
80.0%
80.6%
82.2%
Terminal Bench 2.1
63.8%
56.4%
51.3%
71.3%
82.7%
73.8%
89.5%
MCP Atlas
74.1%
44.7%
64.0%
68.1%
77.8%
78.2%
81.8%
Read this table carefully:
Inkling beats Nemotron 3 Ultra on several agentic rows — consistent with the lab's "broad generalist" story.
Kimi K2.6 and GLM 5.2 still lead Inkling on Terminal Bench 2.1 and several tool-heavy suites — the gap matters if your workflow is unsupervised long-horizon coding without fine-tuning.
Claude Fable 5 leads on HLE text and matches SWE-bench Verified — but Fable remains a permissioned closed model for many teams after the export-control cycle documented in explainx.ai's open-source alternatives map.
Inkling's SWE-bench Verified number uses a bash-only harness; Terminal Bench 2.1 uses an internal coding harness with zero scores assigned to web-search-contaminated solutions. Cross-vendor comparisons still need your repo's 500-ticket eval — not launch tables.
Design Arena and artifact quality
On Design Arena's Agentic Web Dev leaderboard (blinded human head-to-head on generated web apps), Inkling scores ~1257 — top open-weights tier, below Claude Sonnet 5 and Claude Fable 5 but above Kimi K2.6 and GPT 5.6 Sol in the published ordering.
That human-judgment signal matters for teams using models to ship UI artifacts, not just pass unit tests. It pairs with the lab's demos: one-shot resume-filler web apps, nine-page editorial PDFs, and a 40-iteration multiplayer snake game refined with an external reviewer model.
Epistemics: forecasting and censorship resistance
Thinking Machines groups calibration, instruction following, and censorship resistance under "epistemics":
Eval
Inkling
Notes
ForecastBench Brier (no search)
61.1 ± 0.79
Competitive with GPT-5.5 and Gemini 3.1 Pro
ForecastBench Brier (with search)
63.7 ± 0.82
RL trained on proper scoring rules
Cognition Propaganda/Censorship eval
Strong non-compliance
Answers directly on censored topics per Cognition's 2026 eval
For production forecasting agents, calibration often beats peak reasoning — a fine-tuning angle Inkling explicitly targets.
Safety
Benchmark
Inkling
Nemotron 3 Ultra
Kimi K2.6
GLM 5.2
FORTRESS (adversarial)
78.0%
77.6%
65.6%
71.3%
FORTRESS (benign)
95.9%
90.5%
97.2%
90.0%
StrongREJECT
98.6%
98.7%
99.8%
98.5%
Inkling leads open-weights models on FORTRESS adversarial refusal without over-refusing benign analogs — relevant because open weights + Tinker fine-tuning raises classic safety uplift questions. Thinking Machines notes ongoing study of how fine-tuning affects safeguards.
Controllable thinking effort — the cost curve that matters
Single-number benchmark posts hide the binding constraint in production: tokens per successful task.
Inkling's effort sweep (0.2 → 0.99) shows:
Terminal Bench 2.1, Humanity's Last Exam, and IFBench all rise with effort — but diminishing returns kick in above ~0.8 on several curves.
At comparable Terminal Bench scores, Inkling uses about one-third the tokens of Nemotron 3 Ultra.
Practical guidance:
Workload
Suggested effort
Why
Interactive coding in OpenCode
0.4–0.6
Fast iteration; human in the loop catches errors
Overnight agent batch jobs
0.85–0.99
Maximize pass rate; cost amortized
Synthetic data generation
0.3–0.5
Volume matters; grader filters quality
Customer-facing chat
0.5–0.7
Latency SLA + acceptable reasoning depth
Set effort from inside the harness — Thinking Machines randomized tool schemas during training so Inkling is less brittle to harness choice. For local open-model + OpenCode setup patterns, see how to run open-source models locally.
The self-finetuning demo: Inkling → Tinker → OpenCode
The launch's most cited demo is not a benchmark — it is a closed customization loop:
User asks Inkling (inside OpenCode) to fine-tune itself via Tinker into a lipogram that never outputs the letter "e."
Inkling writes objective.py, training YAML, and a self_update.py harness script.
Training runs 96 steps (~27 minutes).
Inkling stages new weights and relaunches — answering "what should I do when my team releases a large language model?" without any "e" characters.
That loop is the manifesto in executable form: the model authors its own post-training job, evaluates against a rubric, and hot-swaps checkpoints — not a human copying JSONL into a notebook.
For teams already on OpenCode as a model-agnostic harness, Inkling is now a first-class story alongside GLM 5.2 and Kimi endpoints. The lab also shipped tml-renderer for reliable sampling with tool calls, reasoning content, and multimodal inputs during post-training.
Inkling-Small preview — when 12B active is enough
Inkling
Inkling-Small (preview)
Active params
41B
12B
HLE (tools)
46.0%
46.6%
SWE-bench Verified
77.6%
77.4%
Terminal Bench 2.1
63.8%
52.7%
SimpleQA Verified
43.9%
20.9%
IFBench
79.8%
83.4%
Inkling-Small is the cost/latency sibling — strong on reasoning and instruction following, weaker on terminal agent tasks and factuality. Thinking Machines is still testing before releasing full weights. Treat the preview as directional, not production-pinned.
Availability, pricing, and deployment paths
Path
Details
Tinker fine-tuning
64K or 256K context; 50% launch discount (limited time); cookbook updated with Inkling-native recipes + audio post-training examples
Playground
Free chat with agentic web search for a limited time — try "feel" before GPU spend
Self-hosting a 975B MoE is not a laptop project. Realistic paths: API for exploration, Tinker for specialization, multi-GPU cluster only after internal eval proves lift over Kimi K2.7 or GLM 5.2 on your tasks.
Honest limitations — what the issue tracker would say if this were GitHub
Not peak agent scores — Kimi K2.6 and GLM 5.2 beat Inkling on Terminal Bench 2.1; GPT 5.6 Sol and Claude Fable 5 lead wider on HLE with tools.
Harness sensitivity — SWE-bench and Terminal Bench numbers depend on harness choice; Thinking Machines uses different harnesses per suite. Compare on Senior SWE-bench methodology before switching CI defaults.
Factuality gap — SimpleQA Verified at 43.9% trails Claude Fable 5 (77.3%) and GPT 5.6 Sol (68.3%). Fine-tune or RAG for customer-facing factual workloads.
975B operational weight — Full self-host is enterprise-GPU territory; most teams should start Tinker/API, not download-and-pray.
Inkling-Small incomplete — Preview only; factuality and terminal scores lag full Inkling.
Fine-tuning safety unknowns — Strong base refusals may shift after customization; budget red-team eval on every Tinker checkpoint you promote.
Bootstrap lineage — Initial SFT used Kimi K2.5 synthetic data; understand license and vendor posture if your compliance team tracks training provenance.
How Inkling fits the July 2026 open-weights landscape
Inkling does not obsolete enterprise open-weight migration playbooks — it extends them with a Tinker-native customization story the Nemotron and Kimi releases did not center as explicitly.
Actionable starting points
Try Inkling in Tinker Playground
Use the Tinker console Inkling Playground before committing fine-tune spend — Thinking Machines added integrated agentic web search for launch.
Fine-tune sketch (conceptual)
yaml
# tinker_cookbook-style supervised run — see Thinking Machines cookbookmodel:thinkingmachines/Inklingcontext_length:65536# or 262144epochs:3objective:org_rubric_grader# your domain rubric + claims grader patterneffort_default:0.6
Pair with claims verification (agentic web search grader) if you are fine-tuning for customer support — mirrors Inkling's RL recipe.
OpenCode + hosted Inkling API
bash
# Install OpenCode if needed — see explainx.ai OpenCode guide
curl -fsSL https://opencode.ai/install | bash
cd your-project && opencode
# /connect → add Together/Fireworks/Baseten Inkling endpoint# /models → select Inkling# Set effort in system prompt or provider options per Thinking Machines docs
Inkling is Thinking Machines Lab shipping the manifesto as weights. It is a 975B multimodal MoE with honest positioning: not the strongest model on every chart, but the most deliberate open-weights foundation for customization — Tinker fine-tuning, controllable thinking cost, encoder-free audio/vision, and a self-finetuning demo that only makes sense if you believe organizations should own specialized models.
Before you rewrite production defaults: run your tickets through Inkling at effort 0.5 and 0.99, compare against Kimi K2.6 and GLM 5.2, and fine-tune one narrow workflow on Tinker. The lab's bet is that post-training lift closes gaps benchmarks show today — your org's data is the only honest test of that claim.
Inkling specifications, benchmarks, and availability reflect Thinking Machines Lab's July 15, 2026 announcement and July 16 social traction estimates. Harness variance, contamination risk, and fine-tuning safety uplift mean enterprise teams should run internal eval before production commitment — vendor tables are orientation, not contracts.
