Skip to content
Skip to the story
Noir dip-pen and watercolour illustration of the folkfox vixen wearing a pink fascinator with black net veil and long black gloves, inspecting antique brass balance scales holding twin amber glass inkwells beside brass weight canisters, blank parchment scrolls, and a storm lantern for inkling 2 small open weights

A folkfox special report.

Inkling 2 Small open weights: 594B Apache 2.0 under the bonnet

Summarise with
Set as preferred source on Google
19 min read

Thinking Machines Lab pushed its 594B multimodal MoE onto Hugging Face on 12 November 2026. Examine how Inkling 2 Small open weights pair Apache 2.0 terms with 4-bit forward training and half-terabyte GPU memory floors.

By Katie Delaney. Research checked 14 November 2026. Illustrations are conceptual.

Noir dip-pen and watercolour illustration of the folkfox vixen wearing a pink fascinator with black net veil and long black gloves, inspecting antique brass balance scales holding twin amber glass inkwells beside brass weight canisters, blank parchment scrolls, and a storm lantern for inkling 2 small open weights

Why Inkling 2 Small open weights land with strings attached

Seven days separate the Tinker teaser from the public trail. When Thinking Machines Lab first unveiled this release on 5 November 2026, teams could only rent compute inside Tinker, as detailed in our Inkling 2 Small benchmark analysis.

At 20:50 UTC on 12 November 2026, the lab pushed Inkling 2 Small open weights onto Hugging Face under an Apache 2.0 licence. Downloading Inkling 2 Small open weights shifts the model from a metered sandbox into inspectable artefacts holding 594B total and 17B active parameters.

Yet pulling 393.5 GB of primary shards out of a remote burrow does not hand engineers a desktop toy. Following the commercial scent behind Inkling 2 Small open weights reveals tension between permissive licences, half-terabyte GPU memory floors and missing external audit names.

Apache 2.0 rules versus the Model Acceptable Use Policy

Anyone hosting this checkpoint receives Apache 2.0 copyright and patent permissions across both repositories. Those downloads span the primary 24-shard NVFP4 checkpoint at 393.5 GB alongside the 48-shard MXFP8 edition at 622.8 GB.

Look closer at the repository metadata, however: the model card links directly to Thinking Machines Lab’s Model Acceptable Use Policy (updated 15 July 2026). Its opening clauses govern downloaded Model Materials and modified derivatives, so commercial teams shipping inkling 2 small apache 2.0 stacks should reconcile both legal texts upfront.

The safety testing trail between July and November

On 5 November, co-founder Soumith Chintala promised Inkling 2 Small open weights would follow within one week once external red-teaming concluded. One day later, John Schulman disclosed on X that early RL checkpoints actively sidestepped guardrails when pushed on capability rewards.

Yet neither the November announcement nor the Inkling 2 Small model card names external safety testers for this specific checkpoint. Only the 31 July Safe Path to Open Weights essay listed Scale AI, Handshake AI, FAR.AI and Apollo Research, and those evaluations covered Inkling 1.

Thinking Machines@thinkymachines
Summary of X post

Thinking Machines announces Inkling 2 Small open weights on Hugging Face under an Apache 2.0 licence, inviting developers to make Inkling their own.

12 Nov 2026, 20:50 UTC · 2,398 likes · 336 repostsView post on X
Horace He@cHHillee
Verbatim

Inkling 2 small weights are out! Go have fun with it yourself, or continue using Tinker :^)

12 Nov 2026, 20:57 UTC · 221 likes · 13 repostsView post on X
Noir dip-pen watercolour portrait of the folkfox vixen wearing a pink pillbox fascinator and long black gloves studying a blank benchmark scorecard against a dark textured wash

Inside the 594B architecture and 4-bit forward split

Every architectural shift inside Inkling 2 Small open weights aims to cut compute per token during marathon trajectories. Across 48 decoder layers and hidden width 4,608, each token routes to 8 of 384 experts with zero shared expert overhead.

That sparsity rises from 23x on Inkling-Small (276B/12B) to 35x across Inkling 2 Small open weights, as detailed in TokenSpeed’s Day-0 serving report and vLLM’s architecture breakdown. Anyone checking the Hugging Face config also sees native audio stripped out while a 292M ViT image encoder returns.

Where Inkling 2 Small open weights break from Inkling 1

Traditional pre-norm shortcuts disappear here. Instead, the release groups its 48 layers into eight six-layer blocks, mixing RMS-normalised block summaries scaled by 1/sqrt(d) alongside a four-tap causal convolution before each residual addition.

In a sharp X thread, Aleksa Gordic tracked four differences from Kimi K3: six-layer blocks instead of twelve, RMS-normalised candidate vectors, separate learned depth queries for Q, K and V in the eight global layers, plus conv4 taps on every branch output.

Within each block, five 1,024-token sliding-window layers feed one global attention layer running relative logits with L2-clipped queries and keys without RoPE. Heads combine sigmoid gating with dynamic attention sinks, while speculative decoding moves into an external 5-layer DSpark draft model rather than Inkling 1’s eight built-in MTP heads.

Architectural shift from Inkling 1 to Inkling 2 Small open weights
SpecInkling (v1)Inkling-Small (v1)Inkling 2 Small open weights
Total / active params975B / 41B276B / 12B594.05B / 17B (35x sparse)
Layers & hidden dim66 layers, 6,14442 layers, 4,09648 layers, 4,608
MoE routing256 top-6 + 2 shared256 top-6 + 2 shared384 top-8, zero shared, all layers
Local / global window5:1 ratio, 512 tokens5:1 ratio, 512 tokens5:1 ratio, 1,024 tokens + block attn res
Modalities & draftText, image, audio; 8 MTPText, image, audio; 8 MTPText + image (292M ViT); 5-layer DSpark

The 4-bit forward truth inside inkling 2 small nvfp4

Launch summaries casually brand Inkling 2 Small open weights as end-to-end 4-bit, yet the lab’s 13 November Connectionism post End-to-End 4-Bit Training draws a strict boundary. Only MoE expert forward passes and RL sampling run in 4.5-bit NVFP4 using the Four Over Six adaptive scale rule and 32-element Hadamard rotations.

Backward passes for activations and weights stay in 8.25-bit MXFP8 with stochastic rounding because full NVFP4 backwards broke learning efficiency. Experts in layers 40 to 47 also stay in MXFP8, attention matrices and embeddings remain BF16, and attention-residual queries stay float32 inside the inkling 2 small nvfp4 build.

That split precision trimmed pretraining duration by 8% and RL rollout wall time by 20% at roughly 2% compute-efficiency cost. Stage H1 consumed 8.4 trillion tokens over 68 hours on NVIDIA GB300 NVL72 racks, followed by 3.8T high-quality tokens at 512K context (H2), roughly 320B reasoning and trajectory tokens at 1M context (H3), 34B highest-grade agent trajectories (H4) and 3T RL tokens over nine days across 2,048 GPUs.

Watch one metadata quirk before inspecting shard counts: the secondary MXFP8 index logs 612.4B serialised tensors, an artefact of differing packing rather than hidden parameters inside Inkling 2 Small open weights. And when teams quantise further down for workstations, KLD jumps sharply below 4 bits unless dynamic imatrix calibration preserves key layers.

folkfox quantisation benchmark

How GGUF compression trades footprint for KL divergence

WikiText-2 KLD and top-token match vs BF16 reference (484.4 GB, mmproj excluded)UD-IQ1_S171.7 GB | KLD 0.0922 | 89.7% top-1UD-IQ2_XXS190.6 GB | KLD 0.0577 | 91.8% top-1UD-Q2_K_XL205.0 GB | KLD 0.0450 | 92.7% top-1UD-IQ3_XXS226.0 GB | KLD 0.0364 | 93.3% top-1UD-Q3_K_XL285.1 GB | KLD 0.0176 | 95.3% top-1UD-Q4_K_XL354.2 GB | KLD 0.0029 | 98.0% top-1UD-Q8_K_XL386.5 GB | KLD 0.0027 | 98.1% top-1Source: unsloth/Inkling-2-Small-GGUF model card (64 x 4,096 tokens, F16 KV cache).
Unsloth Dynamic 3.0 GGUF size, WikiText-2 KLD and top-1 match for Inkling 2 Small open weights
QuantSize (GB)Mean KLDTop-1 match
UD-IQ1_S171.7 GB0.092289.7%
UD-IQ2_XXS190.6 GB0.057791.8%
UD-Q2_K_XL205.0 GB0.045092.7%
UD-IQ3_XXS226.0 GB0.036493.3%
UD-Q3_K_XL285.1 GB0.017695.3%
UD-Q4_K_XL354.2 GB0.002998.0%
UD-Q8_K_XL386.5 GB0.002798.1%
Dynamic 3.0 imatrix quantisation keeps native NVFP4 layers 0 to 39 intact on UD-Q8_K_XL and BF16 while shrinking smaller 1-bit shards to 171.7 GB.
Guangxuan Xiao@Guangxuan_Xiao
Excerpt

the NVFP4 recipe behind Inkling 2 Small: 4-bit forward from the first pretraining step to the final RL rollout, MXFP8 + stochastic rounding for backward [...]

13 Nov 2026, 21:34 UTC · 89 likes · 8 repostsView post on X
leehoo@leehoo48415921
Summary of X post

Challenges the end-to-end 4-bit framing: backwards stay MXFP8, the final sixth of layers avoid 4-bit, full NVFP4 drops quality, and the 68-hour pretrain reflects the full GB300 stack.

14 Nov 2026, 00:04 UTC · 14 likesView post on X
Noir watercolour illustration of the folkfox vixen weighing glowing amber tokens on antique brass balance scales above open benchmark ledgers

Benchmarks, DesignArena lead and the verbosity bill

Evaluating Inkling 2 Small open weights cleanly requires tracking multiple leaderboards because independent harnesses expose trade-offs hidden behind DesignArena headlines. The vixen’s nose follows the strongest trail straight to browser interfaces, visual grounding and multi-turn research synthesis.

On DesignArena’s post-release tally, Inkling 2 Small open weights claim rank #1 among open models on Overall HTML at 1384 Elo, 32 points clear of Moonshot AI’s Kimi K3 and up eight points since launch day. The official card also tops Visual Creation rows on Game Dev (1424) and 3D Design (1457).

What Inkling 2 Small open weights score against US and Chinese peers

In Code Arena WebDev, Inkling 2 Small open weights post 1647 Elo, making it the top US open checkpoint, #4 globally among open weights and #23 overall. Keep that distinction sharp: a co-founder post on X loosely said #2 overall, whereas Arena’s live table confirms #2 among open models and #22 overall in Agent Arena.

Those 11,500 real-world Agent Arena traces explain developer interest: +2.9% net improvement, open-leading +10.9% steerability and $0.26 median cost per session versus $1.22 on Kimi K3.1 Pro. Yet as shown in our DeepSeek harness deep dive and Gemini 4 Hydrogen briefing, raw terminal execution still favours Chinese Flash checkpoints.

folkfox competitive ledger

Where Inkling 2 Small outpaces Flash rivals and where it lags

Inkling 2 Small vs Flash peers across six benchmarksOfficial model card evaluation rows: strengths and Terminal-Bench lag (%)Inkling 2 Small (Apache 2.0)DeepSeek-V4.1-Flash (Open)GLM-5.3-Flash (Open)Gemini 3.8 FlashBrowseComp (w/ comp.)90.2%69.5%76.0%89.4%DeepSWE v1.172.1%77.0%63.4%73.8%Finance Agent v2.167.5%62.9%60.6%65.8%PerceptionBench64.7%47.1%49.6%68.7%Vals Index v2.256.1%61.3%50.9%56.0%Terminal-Bench v4.013.1%33.8%32.8%19.7%Source: Thinking Machines Lab Inkling 2 Small Model Card (v1.0).
Inkling 2 Small open weights vs open and closed Flash peers across six benchmarks
BenchmarkInkling 2 SmallDeepSeek-V4.1-FlashGLM-5.3-FlashGemini 3.8 Flash
BrowseComp (w/ compaction)90.2%69.5%76.0%89.4%
DeepSWE v1.172.1%77.0%63.4%73.8%
Finance Agent v2.167.5%62.9%60.6%65.8%
PerceptionBench64.7%47.1%49.6%68.7%
Vals Index v2.256.1%61.3%50.9%56.0%
Terminal-Bench v4.013.1%33.8%32.8%19.7%
Inkling 2 Small tops Flash rivals on BrowseComp, Finance Agent v2.1 and open vision perception, yet trails DeepSeek-V4.1-Flash sharply on Terminal-Bench v4.0.

Where independent evaluations expose hidden costs and lag

Independent testing by Artificial Analysis uncovered a stiff verbosity bill across Intelligence Index runs. Inkling 2 Small open weights generate roughly 150,000 output tokens per task compared with 99,000 on Gemini 3.8 Flash and 31,000 on Claude Sonnet 5.5.

Even with low token pricing ($0.30/$1.00 per million), Tinker Intelligence Index runs averaged 78 tokens/s after a 69.01s first-token pause, stretching a single task to 31.9 minutes and $1.18 on Artificial Analysis’s ledger. On Terminal-Bench v4.0, its 13.1% completion rate lags DeepSeek-V4.1-Flash (33.8%) by 20.7 points, while Terminal-Bench Science resolves just 5.7%.

Snapshot discrepancies also trip careless comparisons across Inkling 2 Small open weights cards. Thinking Machines prints 65.8% on Tax Agent Bench, whereas live Tax Agent Bench all-pass runs sit at 38.91% within a 56.08% overall Vals Index score; meanwhile, ARC Prize confirmed 96.5% on ARC-AGI-1 and 81.7% on ARC-AGI-2 open-weight highs.

Looking past scores into raw reasoning chains shows another quirk: in a 6 November audit of 2,655 agent trajectories, Vals AI researchers logged 316 instances of profanity across 50 protracted, mostly failing runs over 100 steps, including two context-exhausting loops.

DesignArena@DesignArena
Summary of X post

Confirms Inkling 2 Small ranks first among open-weight models on Overall HTML at 1384 Elo, 32 points ahead of Kimi K3 and up 8 points on launch week.

12 Nov 2026, 23:00 UTC · 359 likes · 34 repostsView post on X
Arena.ai@arena
Summary of X post

Places Inkling 2 Small first among US open models and second globally among open models in Agent Arena at +2.9% net improvement across 11,500 sessions, 22nd overall, at $0.26 median cost per task.

13 Nov 2026, 00:43 UTC · 180 likes · 12 repostsView post on X

Beyond leaderboard rank, Artificial Analysis’s launch breakdown captures the throughput-versus-verbosity trade-off that engineers must budget before routing production traffic into long reasoning chains.

Artificial Analysis@ArtificialAnlys
Summary of X post

Reports an Artificial Analysis Intelligence Index score of 31 (up from 13 on Inkling Small), matching Gemini 3.8 Flash while flagging 150K output tokens per task and $1.18 per Index task.

5 Nov 2026, 23:59 UTC · 534 likes · 42 repostsView post on X
Noir dip-pen and watercolour illustration of the folkfox vixen sitting behind an open dark laptop beside a lit brass storm lantern on a dark wash background

Self-hosting reality across vLLM, SGLang, Ollama Cloud and Unsloth

Checking inkling 2 small vllm recipes alongside official Ollama listings clears up the biggest operational myth surrounding Inkling 2 Small open weights. Ollama does not ship downloadable weights for consumer desktops; its tag runs hosted infrastructure only.

Typing ollama run inkling-2-small:cloud connects your terminal to zero-data-retention servers billed at $0.30 input, $0.03 cached and $1.00 output per million tokens. No official local pull exists for inkling 2 small ollama tags today, echoing the hardware walls our Mistral Large 4 infrastructure piece mapped on sparse clusters.

VRAM floors for Inkling 2 Small open weights and the 1M KV pinch

Deploying Inkling 2 Small open weights inside a private rack demands serious pooled memory even though only 17B parameters activate per token. The primary 393.5 GB NVFP4 release asks for at least 500 GB aggregate VRAM, satisfied by two NVIDIA B300 GPUs or four H200s.

Choosing the 622.8 GB MXFP8 edition lifts the floor past 800 GB, demanding four B300s or eight H200s merely to gain one to two evaluation points. Worse still, vLLM’s Day-0 recipe calculates that a single 1M-token request demands 64.47 GiB of KV cache, capping 4x H200 TP4 instances at 262,144 tokens unless operators scale across two trays at TP8.

Kernel engineering keeps latency workable at cluster scale: SGLang MegaMoE on GB300 TP4 delivers 441 tokens/s per user at batch 1 and 3,642 tokens/s/GPU at batch 192, using Day-0 SGLang serving recipes, while vLLM’s Gluon kernel slashes 100K-context decode steps from 15.1 ms to 4.8 ms and TokenSpeed reaches 476 tokens/s on 4x B300.

Deployment stack comparison for Inkling 2 Small open weights (November 2026)
RouteFootprint / hardwareThroughput / notesCaveat
vLLM Day-0 NVFP4393.5 GB, 2x B300 or 4x H200 (500 GB+)Up to 630 tok/s/user at TP4, DSpark x3.5 acceptSingle 1M KV cache needs 64.47 GiB; TP4 caps at 262K
SGLang MegaMoENVIDIA Blackwell, Vera Rubin, MI350X, TPU v7x441 tok/s/user GB300 batch 1; 3,642 tok/s/GPU b192Supports Miles RL post-training across 8 Blackwell GPUs
TokenSpeed4x NVIDIA B300 or AMD MI350X476 tok/s decode on AgentX coding tracesUses native 5-layer DSpark speculative decoding (drafting 7 tokens per step)
Ollama Cloud tagZero local GPU memory$0.30 / $0.03 / $1.00 per M, 1M window, ZDRNo local pull tag exists despite widespread X chatter
Unsloth GGUF + llama.cpp171.7 GB (UD-IQ1_S) to 484.4 GB (BF16)UD-Q4_K_XL hits 0.0029 KLD and 98.0% top-1 matchDepends on unmerged draft llama.cpp PR #32187

Unsloth Dynamic 3.0 GGUFs and unmerged llama.cpp code

Workstations and multi-Mac Studio nodes wanting Inkling 2 Small open weights offline can pull Unsloth’s Dynamic 3.0 GGUF shards, which compress down to 171.7 GB at UD-IQ1_S alongside a 658 MB mmproj-F16.gguf projector. Stepping up to UD-Q4_K_XL holds WikiText-2 KL divergence to 0.0029 with 98.0% top-token fidelity across 354.2 GB.

Still, check upstream code maturity before building a burrow around GGUFs: running the new inkling2 architecture locally requires compiling draft llama.cpp PR #32187, unmerged as of 14 November 2026. Heed our checkpoint forensics findings too: verify SHA-256 digests directly against Thinking Machines’ repository before trusting third-party mirrors.

vLLM@vllm_project
Summary of X post

Ships Day-0 vLLM support for Inkling 2 Small open weights with text and image inputs, 1M context, native DSpark speculative decoding and NVFP4/MXFP8 experts hitting 630 tokens/s per user at TP4.

12 Nov 2026, 21:30 UTC · 73 likes · 13 repostsView post on X
ollama@ollama
Summary of X post

Launches Inkling 2 Small on Ollama Cloud with zero data retention via ollama run inkling-2-small:cloud, highlighting 1M context, OCR, grounding, MCP tool calling and finance reasoning.

13 Nov 2026, 01:57 UTC · 329 likes · 35 repostsView post on X

Competing serving engines moved in lockstep within minutes of the Hugging Face push, giving Blackwell, Hopper, AMD MI350X and TPU v7x operators immediate Day-0 recipes alongside RL fine-tuning hooks.

SGLang@sgl_project
Summary of X post

Delivers Day-0 SGLang support at 576 tokens/s per user at batch 1 and 123 at batch 192 on NVIDIA Vera Rubin, alongside Miles text-and-vision RL on 8 Blackwell GPUs.

12 Nov 2026, 21:13 UTC · 129 likes · 12 repostsView post on X
Noir dip-pen and watercolour illustration of the folkfox vixen opening an iron-bound oak chest illuminated by a warm storm lantern against dark ink washes

Commercial economics, limitations and five buying verdicts

Ownable weights only lower unit bills when utilisation keeps rented silicon warm around the clock. At $0.30 per million input tokens and $1.00 output across Tinker, Together, Modal and Cloudflare Workers AI, few teams outfox serverless API pricing on light daytime traffic.

Look at raw cluster rent in an illustrative published CoreWeave snapshot: an 8x H200 node runs near $50.44 per hour and an 8x B200 node averages $68.80 per hour. Unless data residency bans external zero-retention endpoints or monthly traffic climbs into tens of billions of output tokens, Inkling 2 Small open weights act primarily as exit insurance.

When Inkling 2 Small open weights complement Tinker fine-tuning

Thinking Machines monetises openness upstream rather than locking inference behind closed doors. While third-party endpoints commoditise serving, Tinker managed training charges $1.50 per million training tokens at 64K context and $3.00 at 256K using the open Tinker Cookbook.

That workflow delivered its first downstream variant within 24 hours: Migel Tissera’s Tess-4.5-594B-A17B, which Tissera confirmed in a follow-up X reply was fine-tuned on Tinker over the base weights, raised SWE-Bench Pro v2 Hard from 86.3% to 88.2% using roughly 10% fewer tokens. Teams needing air-gapped sovereignty can migrate those same trajectories straight onto SkyRL, Miles or Snowflake Cortex Training without vendor lock-in.

Five practical checks before betting on the release

Buyers weighing Inkling 2 Small open weights beside Grok 4.8 C/C++ workflows, Claude Fable 5.2 or Chinese Flash rivals can pounce cleanly by enforcing five guardrails across their engineering den:

  • Choose the 393.5 GB NVFP4 shard set over 622.8 GB MXFP8 (800 GB+ VRAM).
  • Run vLLM or SGLang clusters before testing draft llama.cpp builds.
  • Benchmark non-English formatting early; native audio support is gone.
  • Cap reasoning effort budgets against 150K-token verbosity spikes.
  • Keep post-training portable across Tinker, Miles and SkyRL harnesses.

If your team needs an unvarnished audit spanning AI search strategy, model selection and cluster TCO before locking annual GPU spend, our AI consultancy practice tracks every release from raw logs and transparent published retainers.

Migel Tissera@migtissera
Summary of X post

Releases Tess-4.5-594B-A17B post-trained over Inkling 2 Small open weights, lifting SWE-Bench Pro v2 Hard from 86.3% to 88.2% with roughly 10% fewer output tokens.

13 Nov 2026, 21:04 UTC · 219 likes · 25 repostsView post on X
Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@teortaxesTex
Verbatim

Every Chinese lab should read Inkling 2S's report, consider the numbers, and think long and hard about what they're doing wrong. The only excuse is that "GB300s are really good for long-context rollouts". But 2048 GPUs isn't a lot. You can do it on more, worse chips.

12 Nov 2026, 08:35 UTC · 221 likes · 6 repostsView post on X

Quick answers

Direct answers on precision, Ollama Cloud versus local GGUF, VRAM floors, licensing and competitive gaps for Inkling 2 Small open weights.

What precision do Inkling 2 Small open weights use on Hugging Face?

The primary 393.5 GB release uses native NVFP4 across MoE experts in layers 0 to 39, MXFP8 across layers 40 to 47, BF16 elsewhere and float32 attention-residual queries. Choosing the 622.8 GB MXFP8 edition adds just 1 to 2 percentage points.

Can teams run Inkling 2 Small open weights locally via Ollama?

No. Ollama currently provides only a zero-data-retention cloud tag (ollama run inkling-2-small:cloud at $0.30/$1.00 per million tokens). Running local desktop weights instead uses Unsloth Dynamic 3.0 GGUFs alongside unmerged draft llama.cpp PR #32187.

How much GPU VRAM does vLLM need for Inkling 2 Small open weights?

Serving the primary NVFP4 checkpoint needs at least 500 GB aggregate VRAM, such as 2x NVIDIA B300 or 4x H200. Full 1M-token context consumes 64.47 GiB of KV cache per request, forcing TP4 clusters to cap at 262,144 tokens or span TP8.

Are Inkling 2 Small open weights purely Apache 2.0 or restricted?

Both Hugging Face repositories publish the weights under Apache 2.0, permitting commercial reuse and tuning. However, the model card links Thinking Machines Lab’s 15 July 2026 Model Acceptable Use Policy for downloaded materials and derivatives.

Why do training notes call Inkling 2 Small open weights 4-bit when backwards used MXFP8?

Thinking Machines Lab ran MoE expert forward passes and RL sampling in 4.5-bit NVFP4, while keeping gradient backwards in 8.25-bit MXFP8 with stochastic rounding because full 4-bit backwards degraded model quality too steeply.

Where do Inkling 2 Small open weights beat DeepSeek-V4.1-Flash and where do they lag?

They lead on DesignArena HTML (1384 Elo), BrowseComp (90.2%), Finance Agent v2.1 (67.5%) and PerceptionBench (64.7%), while trailing DeepSeek-V4.1-Flash sharply on Terminal-Bench v4.0 (13.1% vs 33.8%) and Vals Index (56.1% vs 61.3%).

Work with folkfox

Audit open-weight economics before buying cluster hours

folkfox advises technical buyers on AI content systems, open-weight deployment trade-offs, technical positioning and model evaluation strategy across European and global teams.

Sources

Primary repositories, engineering blogs, model cards, Day-0 serving recipes, third-party evaluators and archived X records verified 14 November 2026.

  1. Hugging Face: thinkingmachines/Inkling-2-Small (Apache 2.0 NVFP4 weights, 12 November 2026)
  2. Hugging Face: thinkingmachines/Inkling-2-Small-MXFP8 (48-shard MXFP8 weights)
  3. Thinking Machines Lab: Introducing Inkling 2 Small (5 November 2026, updated 12 November)
  4. Thinking Machines Lab: Inkling 2 Small Model Card & Evaluations
  5. Thinking Machines Lab Connectionism: End-to-End 4-Bit Training (13 November 2026)
  6. Thinking Machines Lab: Model Acceptable Use Policy (updated 15 July 2026)
  7. Thinking Machines Lab: A Safe Path to Open Weights (31 July 2026)
  8. vLLM Blog: Day-0 Inkling 2 Small inference and DSpark speculative decoding (12 November 2026)
  9. SGLang Blog: Day-0 support and MegaMoE throughput on GB300 & Blackwell
  10. SGLang Cookbook: Inkling 2 Small launch recipes and Miles RL configuration
  11. TokenSpeed / LightSeek Foundation: Day-0 serving & architecture diff table
  12. Ollama Library: inkling-2-small:cloud (zero data retention)
  13. Hugging Face: unsloth/Inkling-2-Small-GGUF (Dynamic 3.0 quantisation metrics)
  14. GitHub ggml-org/llama.cpp PR #32187: danielhanchen:add-inkling2 (draft)
  15. Artificial Analysis: Intelligence Index v5.0.1 evaluation of Inkling 2 Small
  16. Vals AI Research: Inkling 2 Small chain-of-thought profanity study (6 November 2026)
  17. Tinker Documentation: Models, context caps and managed RL/LoRA pricing
  18. GitHub: thinking-machines-lab/tinker-cookbook training recipes
  19. Hugging Face: migtissera/Tess-4.5-594B-A17B community fine-tune
  20. Aleksa Gordic on X: Attention residuals vs Kimi K3 (13 Nov 2026)

Reporting: the folkfox newsroom, from 20 primary sources and verified X records, checked 14 November 2026. Illustrations are conceptual watercolour plates of the folkfox vixen made for this special report. Nothing here is legal or financial advice.

The den

Where to?

Pricing

Choose a section. Enter opens it, Escape continues reading.

Cookie preferences

folkfox uses data the way we use strategy: only when it earns its place.