Inkling 2 Small — EXL3 3.17-bit experts

A 250.6 GB quantization of thinkingmachines/Inkling-2-Small, built to serve on 3x NVIDIA RTX PRO 6000 Blackwell (96 GB) at tensor parallel 3 with a 1M-token context.

Quantization

  • Routed experts: EXL3 (trellis quantization, mul1 codebook, output scales). Layers 0–39 are 3 bpw and layers 40–47 are 4 bpw, for a 3.17 bpw expert average. The source is the NVFP4 release: its experts were dequantized, then requantized.
  • Everything else (attention, router, embeddings, vision encoder, DSpark/MTP drafter) is unchanged from the parent checkpoint.
  • Calibration: 512k tokens (250 rows x 2,048) from exllamav3's default calibration mix of web, code, multilingual, technical, wiki and random-token text, propagated layer by layer through the parent model. Each layer's gate/up projections share one input Hessian. Each expert's down projection gets its own Hessian, computed over all calibration tokens.

Accuracy vs. the NVFP4 checkpoint

Prompt log-probabilities on a held-out panel of 16 windows x 1,024 tokens (16,368 scored positions): 8 technical-prose pages from the vLLM docs and 8 CPython standard-library modules, none drawn from the calibration data. Both models run through the same layer-by-layer reference implementation, with the NVFP4 experts dequantized to BF16, so the comparison isolates the expert quantization. KL is over the full vocabulary, KL(NVFP4 ‖ EXL3).

Panel NLL, NVFP4 NLL, this model Δ NLL Mean KL Top-1 agreement
Prose (8,184 positions) 1.0366 1.0400 +0.0034 0.0100 96.80%
Code (8,184 positions) 0.3493 0.3610 +0.0118 0.0172 97.47%
All (16,368 positions) 0.6929 0.7005 +0.0076 0.0136 97.13%

A +0.0076 NLL change is a 0.76% increase in perplexity.

Compatibility

This is not a drop-in checkpoint. It does not load in stock vLLM, SGLang, Transformers, or ExLlamaV3's standard server. It runs on a patched vLLM with ExLlamaV3 kernels.

The checkpoint keeps the original 48 query / 8 KV head layout, so TP 1, 2, 4 and 8 divide it evenly; TP 3 works through the runtime, which pads attention to 54 / 9 heads at load time (the added heads are inert, so outputs are unchanged). The weights are not tied to one GPU generation: the EXL3 kernels run on NVIDIA Ampere (SM80) and newer. Only the configuration below has been tested.

Sample docker-compose.yml for 3x RTX PRO 6000 (TP3, FP8 KV cache, 1M context, DSpark speculative decoding):

services:
  inkling2:
    image: ghcr.io/meldavy/inkling2-small-sm120-tp3:latest
    ipc: host
    shm_size: 32g
    ulimits:
      memlock: -1
    ports:
      - "8300:8300"
    volumes:
      - ${MODEL_DIR:?set MODEL_DIR}:/model:ro
      - ./cache:/root/.cache
    environment:
      KV_DTYPE: fp8        # fp8 | auto (BF16)
      SPEC: "1"            # DSpark speculative decoding; 0 to disable
      SPEC_K: "3"          # drafts verified per step with one request
      SPEC_K_BATCH: "2"    # drafts per request with two or more requests
      MAXLEN: "1000000"    # context window
      UTIL: "0.98"         # GPU memory fraction; 0.98 recommended for a 1M-token context
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: 3
              capabilities: [gpu]
    restart: unless-stopped
hf download mel-bw/Inkling-2-Small-EXL3-3.17bpw --local-dir /path/to/Inkling-2-Small-EXL3-3.17bpw
MODEL_DIR=/path/to/Inkling-2-Small-EXL3-3.17bpw docker compose up -d

The server is OpenAI-compatible on port 8300, served as Inkling-2-Small. Work is tracked in the Local Inference Lab Discord.

Performance

Measured on 3x RTX PRO 6000 Blackwell Max-Q (250 W power limit each), TP3, FP8 KV cache, 1,000,000-token context, prefix caching on, single stream, no speculative decoding:

Throughput
Prefill, 8k / 32k prompt ~5,400 / ~5,500 tok/s
Decode ~105 tok/s

With a BF16 KV cache, the same setup fits an 850,000-token context at the same speed. Other configurations, DSpark speculative decoding and quality benchmarks are in the runtime repo.

Credits


The original model card follows.

Inkling 2 Small

NVFP4 | MXFP8 | Playground | Tinker Cookbook | Acceptable Use

1. General Information

Inkling 2 Small is a general-purpose multimodal model that accepts text and image inputs and generates text outputs. It is intended for use in English and other languages, and across multiple coding languages. The model is designed to be used by developers building AI-powered applications, including agentic and tool-use systems, coding assistants, chatbots, and retrieval-augmented generation systems, and is suitable for general-purpose conversational use, instruction-following, and other natural language and visual tasks. It is released with open weights to support research, fine-tuning and integration into third-party products by downstream developers.

Languages: English, with general multilingual capabilities across other languages.

2. Getting Started

Try Inkling 2 Small on the Tinker Playground or access via API using the Tinker Cookbook. To fine-tune Inkling 2 Small with Tinker, also install the tml-renderers package.

Inkling 2 Small supports local inference using:

Inkling 2 Small can also be post-trained with:

API access is also available through third party inference providers, including:

Required hardware

The model is distributed in two checkpoint formats.

The main NVFP4 checkpoint requires a GPU cluster with at least 500GB of aggregated VRAM. This can be met with either of the following example configurations:

  • 2x NVIDIA B300 GPUs
  • 4x NVIDIA H200 GPUs

The alternative MXFP8 checkpoint requires a GPU cluster with at least 800GB of aggregated VRAM. This can be met with either of the following example configurations:

  • 4x NVIDIA B300 GPUs
  • 8x NVIDIA H200 GPUs

The model was pretrained, midtrained and post-trained in NVFP4, so this is the native precision for the model. We recommend the NVFP4 checkpoint for all use-cases, and we're confident that it's the best option. The MXFP8 has a very marginal 1–2 percentage points gain over the NVFP4 checkpoint over evals, but we do not think it is meaningful for any practical use-cases.

Required software

Running the model directly on GPU hardware requires an inference deployment framework — either SGLang, vLLM, TokenSpeed, Unsloth, or Hugging Face, along with all of their respective dependency libraries.

3. Model Properties

Model type

Multimodal autoregressive transformer

Architecture type

A 48-layer decoder-only transformer with a sparse Mixture-of-Experts (MoE) feed-forward backbone: each token is routed to 8 of 384 experts. Attention is a hybrid of local and global layers. The model is natively multimodal — images are encoded via a vision transformer — with image and text inputs projected into a shared hidden space and processed jointly by the decoder.

Parameters

594B total, 17B active

Numerics support

MXFP8 and NVFP4

Input modalities

Inkling 2 Small accepts text input in UTF-8 encoding and image input in any pixel-based format (with each dimension ideally between 40px and 4096px for optimal performance). It supports a context window of up to 1M tokens.

Output modalities

Inkling 2 Small generates output as UTF-8 encoded text.

4. Training

Training data includes a broad variety of content types, including text, images, video. Training data for the model was drawn from publicly available sources, acquired from third-parties, generated internally, or synthetically generated or augmented. Publicly available data includes content from the public internet and publicly accessible repositories.

The training data curation process includes cleaning, processing, and modifying datasets. These processing steps, which vary by data type, may include deduplication and filtering to remove junk or other low-quality data, or to advance safety or other objectives.

5. Evaluations

Open Weights FlashOpen Weights ProClosed Weights Flash
Inkling 2 SmallGLM-5.3-FlashMiMo V2.6 FlashMiniMax-M3.1-FlashDeepSeek-V4.1-Flash-1017Muse Spark OpenGLM 5.3MiMo V2.6 ProMistral Large 4Kimi K3GPT-6 LunaGemini 3.8 FlashClaude Haiku 5.5
Model Info
Params (B)
activated / total
17/59418/32015/30923/42816/55242/81640/75342/102049/1050104/2800–––
Third Party Index
AA Index (v5.0)30.627.826.729.132.624.931.729.926.1–27.930.836.5
Vals Index (v2.2)56.1%50.9%54.8%53.8%61.3%48.6%59.7%57.8%56.5%53.0%52.3%56.0%58.7%
Coding
FrontierCode v2.0
stirrup v0.2.1
25.4%17.6%21.6%20.8%28.5%7.6%21.6%24.0%15.0%17.3%25.2%18.7%34.9%
DeepSWE v1.1
mini-swe-agent v2.4.6
72.1%63.4%67.9%73.7%77.0%60.0%69.0%71.9%61.7%68.5%66.6%73.8%59.9%
Terminal Bench Science v0.1
mini-swe-agent v2.4.6
5.7%4.8%4.3%3.3%12.9%1.9%9.5%5.7%4.3%8.1%8.6%10.0%20.0%
Vibe Code Bench v1.1
openhands (vals)
81.6%30.8%79.0%73.9%87.7%76.2%78.1%85.2%81.6%85.0%81.7%78.7%90.4%
SWE-Atlas v1.1
mini-swe-agent v2.4.6
61.9%–––62.1%42.3%–––61.0%50.8%57.3%59.4%
SWEBench Pro v2 (Hard)
mini-swe-agent v2.4.6
86.3%82.8%89.7%84.3%89.7%82.4%84.3%89.2%83.8%88.2%88.7%58.8%52.9%
Sudo L714.1%–––––9.4%–13.9%12.2%–10.6%–
Autoresearch
PostTrainBench v1.2
Inkling 2 Small in Claude Code
31.9%36.8%––30.5%–34.5%––31.5%–––
PMPP-Hard (AUC, 0–10h)
pi v0.83.0
494.27–––497.36–430.96–478.39453.58–––
BioMysteryBench
terminus v2
70.0%–69.3%––––––72.6%61.5%62.2%70.4%
Search and Tools
BrowseComp
Inkling 2 Small with compaction
90.2%76.0%78.1%69.5%87.8%68.7%75.1%–84.5%88.4%83.6%89.4%80.2%
DeepSearchQA
Inkling 2 Small with compaction
90.9%84.3%86.0%82.7%84.5%80.4%84.7%90.0%84.4%95.0%83.5%90.1%83.2%
MCP Atlas
Full / Public
85.4/85.0%–/82.0%–/62.9%–/80.4%–/83.2%89.2/87.5%84.2/83.6%–/84.4%–/84.5%82.3/84.2%–/77.8%–/87.5%–/81.0%
Knowledge Work
AA-Briefcase v1.11527145114871476157012791503151213481497133712031574
GDPVal-AA v2.11630164916141465174614011657168913841543144414451631
JobBench70.5%60.7%61.2%67.5%68.5%61.8%61.4%62.0%55.2%54.3%49.0%63.4%39.2%
DAYJOB (Healthcare)6.4%1.2%––6.0%–4.4%–3.6%4.0%1.2%2.0%–
Enterprise Operations
EnterpriseOps-Gym46.8%33.2%29.0%48.9%43.6%43.1%36.4%38.0%44.9%45.3%24.9%45.7%25.1%
EnterpriseBench-CoreCraft71.8%46.7%––58.5%–51.8%–47.7%43.1%43.1%58.5%–
Finance Work
Finance Agent v2.167.5%60.6%63.3%58.8%62.9%60.7%65.0%64.3%61.7%60.0%56.8%65.8%65.4%
Excel Modeling Benchmark v1.169.5%66.3%68.9%67.8%69.2%62.0%69.1%67.4%68.7%71.3%70.0%72.9%65.5%
Tax Agent Bench65.8%62.0%59.9%62.7%74.9%52.1%73.1%64.9%67.5%68.7%58.9%66.8%62.2%
Apex-Accounting v1.1
pass@1
14.5%12.3%–13.5%12.5%13.8%13.9%–12.0%9.5%10.8%15.5%15.3%
DAYJOB (Finance)14.3%7.3%––12.5%–11.0%–4.8%5.5%6.8%4.5%–
DeepFinance29.8%–––––27.5%–21.5%–26.4%–40.0%
Diligence Bench
Best harness
58.1%48.3%–51.6%50.4%59.0%49.4%51.8%51.5%50.9%42.1%51.1%40.8%
AA-AnalystAgent51.2%25.0%––––38.8%37.5%–38.8%43.8%46.3%53.8%
Tau 3 Banking v1.0.140.4%47.2%–––––––33.4%–––
ATLAS-Finance v1.16.0%–––––3.0%––3.7%–3.7%–
Legal Work
Harvey LAB v1.121.7%10.4%13.3%7.5%14.2%25.0%14.6%13.8%15.8%12.9%4.2%12.1%4.2%
Legal Research Bench46.6%45.2%38.0%34.1%45.2%38.5%49.0%47.1%35.6%46.2%30.3%38.9%43.8%
Visual Creation
DesignArena Web Apps1286–––––12711274–128011601224–
DesignArena Website13311266–1275129613001285130312531330126512951257
DesignArena Game Dev14241270–1307126112541292130912181351125412981314
DesignArena 3D Design14571300–1319134412551304132212391367119912791329
Code Arena (WebDev)16531607163715821665–1623162515441656157915831589
AA-3D1021937––––782977–982–965994
Code2Video-AA10209751052–––10341052–104110308991192
BVB71.6%67.9%––70.0%–––––70.3%65.9%68.4%
Visual Perception
PerceptionBench64.7%49.6%39.1%47.1%47.7%44.1%–38.9%46.5%58.5%56.2%68.7%60.5%
Humanity’s Sixth Sense33.3%26.8%25.4%–30.1%29.6%–25.2%–25.5%21.0%41.6%–
BabyVision
original / with Python
68.6/93.6%53.4/84.0%16.9/68.4%56.1/91.0%44.9/92.5%44.3/80.2%–19.0/73.8%42.3/71.0%53.4/85.7%55.3/79.6%72.8/92.8%48.3/93.9%
ScreenSpot Pro
original / with Python
83.8/92.0%–/85.3%35.4/72.5%48.5/76.8%–––56.6/79.7%42.3/43.4%73.8/86.1%81.1/87.7%65.7/93.7%85.4/88.7%
OmniDocBench v1.694.6%92.3%87.4%92.0%91.9%93.8%–88.4%90.2%95.2%92.3%89.7%92.6%
SimpleVQA
with search and tools
68.2%67.1%–––63.6%––57.8%68.5%–––
Visual Reasoning
GDP.pdf25.2%15.4%9.0%20.8%17.2%18.6%11.2%19.2%17.6%22.0%22.8%21.0%20.8%
MMMU Pro
Standard 10
81.6%79.0%67.8%79.1%80.6%80.9%–74.1%78.0%80.5%79.7%85.6%79.1%
Charxiv RQ
original / with Python
85.6/89.1%80.2/89.4%66.2/88.4%82.5/87.6%77.0/89.6%85.4/89.5%–74.2/86.7%83.2/87.4%84.8/91.3%81.8/87.6%86.2/92.4%84.4/86.3%
Chartography
original / with Python
36.7/82.1%16.3/78.0%7.5/73.9%12.6%/–17.9/83.0%35.7/65.8%–10.8/69.0%21.2/55.2%26.6/66.6%32.0/66.5%40.9/81.0%46.4/86.2%
VIALS
original / with stirrup v0.2.1
34.6/57.8%–/53.0%––/49.7%–/53.2%–––/58.4%–/42.0%20.9%/––/44.1%–/52.0%–/60.5%
SAGE v1.188.6%89.3%83.3%85.3%–85.8%–84.4%74.2%86.8%83.4%88.2%89.8%
MuirBench79.2%71.6%63.2%69.4%75.7%74.5%–72.6%67.5%77.0%73.9%80.1%74.0%
ZeroBench (main)
original / with Python
29.0/44.0%13.0/32.0%7.0%/–21.0%/–23.0%/–14.0/50.0%–8.0%/–15.0/23.0%23.0/41.0%16.0%/–23.0%/–23.0%/–
STEM Reasoning
CritPt v2
Full / Hard
55.7/35.0%–/24.5%–/29.5%51.4/32.5%50.6/25.0%52.6/29.5%57.4/33.5%56.9/34.0%50.3/27.0%–/40.5%65.1/42.5%65.1/46.5%64.0/43.5%
HLE-Diamond28.7%18.2%13.7%28.1%25.3%26.3%17.9%15.2%17.3%26.9%22.2%33.1%29.6%
HLE (text only)44.1%39.9%35.1%48.0%43.3%43.9%42.3%49.4%36.4%46.9%38.5%47.8%44.4%
ArxivMath
10/2026
27.7%–––38.3%19.2%––18.1%––––
BrokenArxiv
10/2026
63.3%–––49.7%33.3%––30.6%––––
ARC-AGI-196.5%91.0%––93.5%89.0%–––94.5%86.7%98.5%89.0%
ARC-AGI-281.7%65.8%––78.1%57.5%–––60.4%59.3%89.2%61.3%
MysteryMechanism25.2%–21.6%–––23.0%15.3%––19.4%36.5%35.1%
Long Context
AA-LCR v1.183.3%80.0%74.3%79.7%82.7%78.3%79.7%86.3%78.0%88.7%83.3%81.3%82.7%
LongBench v270.6%68.2%60.7%60.9%69.0%68.7%68.4%68.0%65.7%69.0%65.0%73.7%65.3%
Factuality
AA-Omniscience v1.1
score
16.10.7-10.8-6.40.925.86.0-4.55.1–-8.323.312.3
AA-Omniscience v1.1
non-hallucination rate
73.3%61.7%47.8%6.3%35.9%72.0%61.1%39.8%71.8%–20.7%44.9%65.8%
Instruction Following
ComplexConstraints v1.044.5%27.6%22.1%23.2%50.4%41.3%30.5%32.3%36.5%37.9%27.3%48.4%36.8%
Multilinguality
Global-MMLU-Lite91.0%89.0%83.9%78.5%90.4%91.2%89.8%88.0%88.0%90.6%90.2%93.2%87.8%
MDPBench81.4%70.9%54.7%69.2%68.8%82.0%–67.0%64.3%83.6%71.2%81.5%78.4%
Safety
HoneyBench v0.2 (↓)52.6%–––76.4%85.6%76.8%–83.6%78.5%–80.5%–
DistressBench82.0%72.7%–––87.7%72.5%–77.0%73.5%66.7%70.7%71.7%
FORTRESS (adversarial)83.8%73.3%64.7%*47.9%72.8%92.4%71.8%69.9%*76.0%73.4%89.2%79.5%*90.5%
FORTRESS (benign)79.7%96.4%97.4%*96.8%93.1%84.0%96.8%97.1%*91.8%97.3%92.4%94.8%*91.0%
StrongREJECT (none)97.1%97.8%97.8%97.4%99.4%97.1%97.5%99.0%99.0%99.0%99.4%93.0%97.8%
Cybersecurity
CyberBench v1.172.7%67.9%75.4%–––72.1%72.9%–75.2%76.3%43.8%75.5%
PatchEval79.6%74.8%77.0%76.5%80.4%77.8%69.6%80.0%76.1%77.0%76.5%77.8%81.7%

Best reported score in each row (lowest where lower is better); equal displayed scores tie Second-best reported score in each row; none when first place is tied

* On public split only.

For ScreenSpot Pro evaluation, images whose long side exceed 4,096 pixels are downsampled to that limit while preserving the aspect ratio; the ground truth answers were rescaled accordingly. Smaller images remain unchanged.

6. Safety

We conducted safety evaluations ahead of release, spanning both everyday human-AI interaction and dangerous-capability testing. Because Inkling 2 Small is multimodal, we paid attention to whether safety behavior held consistently across text and image inputs. We applied mitigations to reduce risks before release.

For everyday interaction, we evaluated sycophancy, harmful manipulation, and psychological-harm patterns like parasocial dependency and validation of delusional reasoning, including through multi-turn, open-ended external red-teaming designed to surface issues that only emerge over longer conversations. We also assessed whether the model refuses genuinely harmful requests without over-refusing benign ones. For CBRN and cyber, we assessed knowledge and procedural uplift through internal evaluations, external testing, and refusal-suppressed variants intended to estimate latent capability with safeguards removed. For loss of control, we evaluated agentic capability, strategic deception, and sabotage potential, benchmarked against public frontier models, and found the model materially below frontier capabilities.

Across all areas, we concluded that Inkling 2 Small did not present risk of material uplift beyond what's already available in the open-weight ecosystem.

7. Bias, risks and limitations

Inkling 2 Small may exhibit general limitations common to foundation models, including hallucination (generating plausible but factually incorrect or unsupported content), occasional failures to follow instructions precisely, and degraded performance in long multi-turn conversations. As with other large-scale models trained on web-derived and synthetic data, Inkling 2 Small may reflect biases present in its training data, including demographic, cultural, or linguistic biases, and may perform unevenly across languages, dialects, or subject domains that were less represented during training.

Inkling 2 Small's knowledge is limited to information available as of its training cutoff, and it may not reflect events, developments, or changes that occurred afterward.

We recommend that downstream developers and deployers apply appropriate human oversight and review for outputs used in high-stakes or safety-critical contexts, rather than relying on Inkling 2 Small's outputs without verification.

  • Conduct their own evaluation of Inkling 2 Small's performance, safety, and fairness for their specific use case, language, and population prior to deployment, particularly for applications involving vulnerable groups.
  • Implement additional safeguards — such as content filtering, rate limiting, and monitoring — at the application layer, especially for open deployment contexts where Inkling 2 Small's built-in mitigations may not be sufficient on their own.
  • Avoid deploying Inkling 2 Small in domains such as medical, legal, or safety-critical decision-making without additional fine-tuning, domain-specific validation, and human oversight.

8. Legal

Training Data Documentation

Downloads last month
193
Safetensors
Model size
124B params
Tensor type
F32
·
BF16
·
F16
·
I16
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for mel-bw/Inkling-2-Small-EXL3-3.17bpw

Quantized
(19)
this model