Flint NVFP4 · azureml v1

# Flint NVFP4 for local inference ## Overview Flint is a mixture-of-experts language model from Recursive, built for coding agents. It is designed to investigate codebases, fix bugs, implement featur…

azureFoundry/azureml/Flint-NVFP4@1

Flint NVFP4 for local inference

Overview

Flint is a mixture-of-experts language model from Recursive, built for coding agents. It is designed to investigate codebases, fix bugs, implement features, and write tests, and it supports reasoning and multi-turn tool use.

This variant provides Flint as NVFP4 GGUF weights, primarily intended for local inference on NVIDIA GPUs using llama.cpp with CUDA. The 22.9 GB model file includes the weights, tokenizer, and chat template.

Model details

PropertyDetails
ModelFlint
ProviderRecursive
Model typeDecoder-only mixture-of-experts Transformer
Total parameters39.8B
Active parametersApproximately 2.5B per token
Primary useCoding agents
DeploymentLocal inference
Context lengthUp to 262,144 tokens, shared by prompt and output
Input / outputText / text
QuantizationNVFP4 (GGUF)
Model size22.9 GB
Knowledge cutoffJune 2026

Capabilities

Flint is designed for agentic software development tasks, including:

  • Codebase investigation
  • Bug fixing
  • Feature implementation
  • Test writing
  • Reasoning and multi-turn tool use

Local inference

Flint NVFP4 runs on llama.cpp built with CUDA, using a release with Flint support. At least 32 GB of GPU memory is required, or unified memory on Spark systems.

The recommended generation settings are temperature 1.0, top-p 0.95, and top-k 20. Use the bundled chat template and preserve returned reasoning content when replaying conversation history. Reasoning counts toward the output budget, so allow at least 32,768 output tokens for hard problems.

For more details, see the Flint NVFP4 GGUF model card and the Flint model card.

Training

Flint was trained from scratch through pretraining, mid-training, supervised fine-tuning, and reinforcement learning. Training data included publicly available and crawled web documents, text and code, commercially licensed data, and AI-generated synthetic data.

Limitations and responsible use

Flint can produce incorrect or harmful content, faulty code, or inappropriate tool calls. Test generated code and review consequential changes before applying them. Tool-using applications should enforce their own permissions and execution boundaries.

Long reasoning can increase latency and exhaust the output budget before an answer is complete. The 262,144-token context is a capacity limit, not a guarantee of uniform accuracy across all sequence lengths or tasks.

License

Flint is released under the Apache License 2.0.

Specifications & capabilities

Input
text
Output
text
Context window · tokens
262144
Max input · tokens
229376
Max output · tokens
32768
Knowledge cutoff
2026-06
Capabilities
toolCalling

Provider specifications & access

1–1 / 1

Flint NVFP4 · azureml v1 ↗

source model id
azureml/Flint-NVFP4@1
hosting
foundry-managed-compute
platforms
  • Microsoft Foundry
license
apache-2.0
context length
262144
max input tokens
229376
max output tokens
32768
architecture
input_modalities
  • text
output_modalities
  • text
capabilities
  • toolCalling

Source specification observations

1–4 / 4

azureFoundry ↗

field
context_length
value
262144
normalized value
262144
scope
source-model
source id
azureFoundry
confidence
1

azureFoundry ↗

field
max_input_tokens
value
229376
normalized value
229376
scope
source-model
source id
azureFoundry
confidence
1

azureFoundry ↗

field
max_output_tokens
value
32768
normalized value
32768
scope
source-model
source id
azureFoundry
confidence
1

azureFoundry ↗

field
knowledge_cutoff
reported field
training_data_cutoff
value
June 2026
normalized value
2026-06
precision
month
scope
source-model
source id
azureFoundry
confidence
1

Original sources

1–3 / 3

Microsoft Foundry · Model page ↗

Microsoft Foundry · Model documentation ↗

Microsoft Foundry · Catalog ↗

Source links

1–1 / 1

Explore more models

Models

Loading
EVIDENCE / COMPARISON

Benchmarks

Explore every benchmark

Loading

These charts contain model-linked results. The charts above also include unlinked source profiles. Compare scores within each evaluation. Model quality, usage, transparency and system performance measure different things.

Install TensorTrace