ENPT

MISA AI 2.0 Decision Plus

Last updated:

Open—$0/M input

Quadrant scores

See the full quadrant

Scored with the public quadrant rubric (maturity × capability; bubble = adoption). Revise as evidence lands.

  • Maturity5.1/10

    Apache-2.0 weights (Gemma-4-31B-it fine-tune, 62.6 GB bf16), a card with SGLang serving of /v1/systemone and /v1/decisions, tips and limits. Company-run (MISA JSC) but no SLA, no hosted API, no integrations found.

  • Capability5.6/10

    Noul, choice and score over text and images, English and Vietnamese. No benchmark, ECE or latency number on the card; not on JevBench, DI 0.3 or AA. Calibration is only a tag. Scored on declared types, with accuracy and calibration at low defaults.

  • Adoption1/100

    Published 16 Oct: 11 downloads, 0 likes, no launch post, no integrations or users found.

Vendor claims

publicado (OK do Fabricio 17/10 BRT).

Read this first. The card publishes no benchmark, calibration or latency numbers, and the model is on no board. The scores below rest on documented features only.

MISA AI 2.0 Decision Plus is a decision model from MISA JSC (the card gives no further description of the company), fine-tuned from Gemma 4 31B-it. It answers structured questions about text or images with probabilities instead of generated text, reading the state once in a single prefill pass.

Specs

Attribute Value
Company MISA JSC
Published 16 Oct 2026 (HF repo created 08:21 UTC, 05:21 BRT)
Weights misa-jsc/misa-ai-2.0-decision-plus, bf16 safetensors in two shards (~62.6 GB), Gemma4ForConditionalGeneration, ~31B dense, 60 layers
Licence Apache-2.0 tag; the card says the weights follow the Gemma 4 licence (Apache 2.0), inherited from the base
Languages English and Vietnamese
Input Text and images; 256K context declared
Question types noul (yes/no), choice (one label with per-option probabilities and a confidence), score (ordinal levels). Each answer also returns a “label mass” (the share of probability on valid labels) as a fit check
Serving “Recent SGLang releases serve /v1/systemone and /v1/decisions out of the box” (--default-chat-template-kwargs '{"enable_thinking": false}')
Price Self-hosted. No hosted API found
Benchmarks None published

Relation to linne-oak

linne-oak is another Gemma 4 31B, English and Vietnamese, Apache-2.0 decision model published five days earlier with the same style of tags, and its card describes it as an independent personal project. The MISA card does not mention linne-oak, and nothing public links the two. No claim is made here beyond that observation.

Fit / anti-fit

Fit when you need: an open 31B decision model for Vietnamese and English text or images, in the /v1/systemone shape, with a “label mass” signal for ill-fitting questions.

Anti-fit when you need: any measured accuracy or calibration, a hosted API, or small hardware (62.6 GB of weights).

Limits

  • No numbers anywhere: no board, no ECE, no latency, no training data description.
  • The card’s own caveat: probabilities are calibrated on the training distribution; check calibration on very different domains and set safety thresholds on your own traffic.

Why it is in the catalog

Open Apache-2.0 weights, a typed /v1/systemone surface and company authorship meet the inclusion rule, but evidence is thin, so it is scored low. Revisit when a board or the vendor publishes results.

Score working (17 Oct 2026)

  • Maturity = availability 8 × 0.30 + docs 6 × 0.25 + integrations 3 × 0.25 + support 2 × 0.20 = 2.40 + 1.50 + 0.75 + 0.40 = 5.05 → 5.1
  • Capability = types 10 × 0.30 + calibration 4 × 0.25 + latency 4 × 0.25 + accuracy 3 × 0.20 = 3.00 + 1.00 + 1.00 + 0.60 = 5.60 → 5.6
  • Adoption = engagement 1 × 0.35 + ecosystem 2 × 0.35 + production 0 × 0.30 = 0.35 + 0.70 + 0.00 = 1.05 → 1

Availability 8: Apache-2.0 weights on the Hub (as linne-oak). Docs 6: a card with usage, serving and limits but no eval. Integrations 3: SGLang serves /v1/systemone (compatibility = 3). Support 2: a company author, but no SLA or hosted service.

Calibration 4: tagged “calibrated” with no method (the linne-oak precedent). Latency 4: no number (the “no number = 4” band). Accuracy 3: no benchmark or claim at all, below the 6 given to vendor runs.