Thinking Machines Inkling 7/15 one-day triple: ~1T multimodal open-source + vLLM day-0 + SGLang day-0, 380 tok/s/user MTP8 on Blackwell

4 sources

On 2026-07-15 Thinking Machines published Inkling on Hugging Face — a ~1T parameter open multimodal model (image / text / audio native inputs, 45T training tokens, 975B total / 41B active MoE, 1M context, relative attention, 5:1 sliding-to-global hybrid attention, hierarchical MLP patchifier + mel-spectrogram discretization); same day vLLM blog posted day-0 support (NVFP4 + BF16 dual-variant full feature parity, 8 MTP heads, 380 tok/s/user MTP8 on 4x GB200 with mean acceptance length 4.5, 140 tok/s/user without MTP, 1M context accuracy-by-length-bucket 99.09% / 95.68% / 81.36%); same day LMSYS posted SGLang day-0 support (71.7k tok/s input throughput on Blackwell, SGLang Cookbook publishes TP4 / TP8 / H200 / AMD MI350X / MI355X deployment recipes). The three releases together form 'one of the strongest single days in the July open-source AI ecosystem: a 1T multimodal open model + same-day inference engine readiness.'

Why now

7/15 was the strongest single open-source day in early July — a ~1T multimodal open release + vLLM + SGLang all on the same day, turning 'a 1T multimodal model in the repo' into 'runnable on Blackwell today.' Creators can frame this as 'open-source big models aren't shelf-ware — they're same-day inference-ready'.

Why it is worth publishing

Huge demo surface: load thinkingmachines/Inkling BF16 weights from HF, bring up a vLLM endpoint with --model-impl, run a multimodal inference (image + text + audio mixed input), and show the 1M-context accuracy-by-length-bucket data; also benchmark vLLM against SGLang on the same weights.

Evidence basis

Three same-day day-0 launches carry their own heat, and Inkling's self-reported scores (HLE 29.7%, AIME 2026 97.1%, VoiceBench 91.4%, SWE-Bench Verified 77.6%) are model-card-tier numbers — heat is medium-to-high as a single-flagship event but driven by official self-reporting rather than independent reproduction.

“On 7/15 the open-source AI ecosystem turned 'a 1T multimodal model in the repo' into 'runnable on Blackwell today' — Thinking Machines shipped the model, vLLM and SGLang both shipped day-0 support at the same time.”

Angle

Frame the 7/15 Inkling one-day triple (Thinking Machines open-source + vLLM day-0 + LMSYS SGLang day-0) as 'one of the strongest single days in the July open-source AI ecosystem: a 1T multimodal model + same-day inference engine readiness' — bundle the 'repo weights + inference engine + deployment recipes all available the same day' into one piece rather than reading any single benchmark in isolation.

Format

Long-form explainer

Demo idea

Record a 15-minute three-segment demo: 5 minutes loading thinkingmachines/Inkling BF16 from HF and bringing up a vLLM endpoint with --model-impl, running multimodal inference (image + text + audio mixed input) and showing the accuracy-by-length-bucket data; 5 minutes bringing up the same weights under SGLang and comparing throughput; 5 minutes on 'open-source weights + same-day day-0 inference engine' meaning creators don't wait 2-4 weeks for adapter work.

Platform notes

Inkling's headline scores (HLE 29.7% / AIME 2026 97.1% / VoiceBench 91.4% / SWE-Bench Verified 77.6%) are Thinking Machines self-reported (medium risk) — don't paraphrase as third-party validation; the license is not stated in the captured summary (medium risk) — don't guess 'Apache-2.0' or 'MIT'; the 81.36% (586K-805K) accuracy-by-length-bucket figure is vLLM self-reported (medium risk) — don't fill in numbers from memory.

Usable claims

Hugging Face's July 15, 2026 blog post announces Inkling by Thinking Machines as a ~1T parameter open multimodal model that natively accepts image, text, and audio inputs, trained on 45T tokens; the architecture is decoder-only multimodal MoE, 975B total / 41B active parameters, 256 experts with top-6 + 2 shared experts, 1M context window, relative attention (no RoPE), hybrid 5:1 sliding-window-to-global attention, short 1D convolution over hidden states, hierarchical MLP patchifier for vision, mel-spectrogram discretization for audio; variants are Inkling BF16 (2 TB VRAM) and Inkling NVFP4 (600 GB), with Inkling-Small BF16 600 GB (276B / 12B active) / NVFP4 180 GB; self-reported benchmarks for Inkling vs Inkling-Small include HLE text-only 29.7% / 31.6%, HLE with tools 46.0% / 47.8%, AIME 2026 97.1% / 95.5%, GPQA Diamond 87.2% / 89.5%, SWE-Bench Verified 77.6% / 80.2%, SWE-Bench Pro 54.3% / 55.9%, Terminal Bench 2.1 63.8 / 64.69, MCP Atlas 74.1% / 79.2%, MMMU Pro 73.3% / 74.0%, VoiceBench 91.4% / 90.1%; license is not stated in the captured summary.
vLLM and SGLang both shipped day-0 support for TML Inkling on 2026-07-15. vLLM documents full feature parity across NVFP4 and BF16 variants, 8 MTP heads for speculative decoding (up to 9 tokens per forward step), native support for 1M tokens, and quoted throughput on 4x NVIDIA GB200 of 380 tok/s/user with MTP8 (mean acceptance length 4.5) and 140 tok/s/user without MTP; vLLM also documents accuracy-by-length-bucket figures of 99.09% (436/440) at 2K-221K, 95.68% (421/440) at 294K-513K, and 81.36% (358/440) at 586K-805K. SGLang documents day-0 support for the 975B multimodal MoE with 1M-token context, hitting up to 71.7k tok/s input throughput on Blackwell.

Evidence pipeline

From the news

Breakdown

Reading any single Inkling fact point (HLE 29.7% / VoiceBench 91.4% / SWE-Bench Verified 77.6% / 380 tok/s/user / 71.7k tok/s input) in isolation turns into 'benchmark reading.' This piece explains how to use the 'one of the strongest single days in the July open-source AI ecosystem: a 1T multimodal model + same-day inference engine readiness' frame — bundle the 'repo weights + inference engine + deployment recipes all available the same day' into a three-segment comparison so creators can produce 'open-source big models aren't shelf-ware — they're same-day inference-ready' content rather than benchmark reading.

Sources

Risks

Pin links to each source; quote only what the captured summary states; do not paraphrase specific benchmark numbers, performance metrics, license details, paper claims, or architectural details beyond what is stated.
Pin the link to the HF blog post; quote only what the post states; do not paraphrase the headline scores as third-party validation; flag that the license is not stated in the captured summary and that the benchmark numbers are vendor self-reported.

Demo ideas

Load thinkingmachines/Inkling BF16 from HF (2 TB VRAM), bring up a vLLM endpoint with --model-impl, and run a multimodal inference (image + text + audio mixed input) showing the 1M-context accuracy-by-length-bucket data.
Benchmark vLLM vs SGLang on the same Inkling weights on Blackwell (380 tok/s/user MTP8 vs 71.7k tok/s input) and walk through 'same weights, different inference engine' trade-offs.
Walk through the SGLang Cookbook deployment recipes on TP4 / TP8 / H200 / AMD MI350X / MI355X to show 'the same weights deploy on five hardware targets.'