DeepCoder Explained: Microsoft's 2017 Research vs Agentica's DeepCoder-14B, Worth Running in 2026?

There are two unrelated AI projects called DeepCoder. If you want a coding model to download, you want DeepCoder-14B-Preview, an open-source model that Agentica and Together AI announced on April 8, 2025, under the MIT licence. The older DeepCoder is a 2017 research paper from Microsoft Research and the University of Cambridge that used a neural network to guide a search for very short programs.

In short, the 2025 model is a 14.8-billion-parameter reasoning model, trained with reinforcement learning from DeepSeek-R1-Distill-Qwen-14B, and reported at 60.6% on LiveCodeBench v5, 0.3 points below OpenAI's o3-mini at its low reasoning setting. A year after its release there is no newer DeepCoder, and it is still labelled a preview. I think it is worth a test on your own problems, not a default. The 14B figures below are reported by Agentica and Together AI. I found no independent evaluation, and I have not run the model.

Which DeepCoder? Two projects, eight years apart

	DeepCoder (2017)	DeepCoder-14B-Preview (2025)
Made by	Matej Balog (University of Cambridge) and four Microsoft Research co-authors	Agentica and Together AI
What it is	A method, described in a paper	Open model weights, MIT licence
What it does	Predicts which of 34 built-in functions a program uses, then searches for programs of up to five calls	Reasons step by step, then writes a full solution to a programming problem

Sources: Balog et al., ICLR 2017 and Together AI, April 8, 2025.

A third name adds to the confusion. DeepSeek Coder is a separate family of code models from DeepSeek, released from late 2023 in sizes from 1.3B to 33B, with a V2 in June 2024. DeepCoder-14B is not one of them, though DeepSeek is in its lineage: its base model, DeepSeek-R1-Distill-Qwen-14B, is Qwen2.5-14B fine-tuned by DeepSeek on samples curated with DeepSeek-R1.

What Microsoft's DeepCoder did in 2017

The paper, "DeepCoder: Learning to Write Programs" (ICLR 2017), is by Matej Balog, Alexander Gaunt, Marc Brockschmidt, Sebastian Nowozin and Daniel Tarlow. Balog did the work as a Microsoft Research intern. The task is programming by example: given five inputs and their outputs, find a program that maps one to the other. Programs use a small language, loosely modelled on SQL and LINQ, over integers and integer lists: functions such as SORT, FILTER, MAP and SUM, plus a fixed menu of lambdas, 34 elements in all. A neural network reads the examples and predicts how likely each element is to appear. Those probabilities decide which candidates existing search tools, such as the Sketch solver, try first. The training programs were generated from the language itself.

In the main experiment, with test programs five calls long (a search space of about ten billion programs), the guided "Sort and add" search solved the first 20% of 100 problems in 9 seconds instead of 8,181, about 907 times faster. The gain shrank on harder problems, to more than 2 times at 60% solved, and the authors were plain that this covered "only the simplest problems on programming competition websites".

ScienceAlert, crediting New Scientist, reported that DeepCoder could steal code from other programs, and other outlets said the same. My own post that March, Is DeepCoder the End of Human Programmers?, said it "borrows" code from existing software. In 2017 I described it the way it was widely reported; the paper describes something narrower. Nothing is copied, as Quartz reported at the time: the network ranks 34 built-in functions so that a search reaches a five-line program sooner.

What is DeepCoder-14B? Agentica's open coding model

The model card places Agentica's work within UC Berkeley's Sky Computing Lab and Berkeley AI Research. DeepCoder-14B-Preview adds reinforcement learning to DeepSeek-R1-Distill-Qwen-14B, using about 24,000 problems: 7,500 from TACO Verified, 16,000 from PrimeIntellect's SYNTHETIC-1 and 600 from LiveCodeBench, dated May 2023 to July 2024. Training took two and a half weeks on 32 H100 GPUs.

The method is GRPO+, a modified GRPO that drops the entropy loss, filters problems by difficulty offline, and takes three changes from DAPO: no KL loss, a higher clipping bound and no loss on truncated answers. Iterative context lengthening trains at 16K tokens, then 32K, with evaluation at 64K. The dataset, code and training logs are public. A 1.5B sibling, trained from DeepSeek-R1-Distill-Qwen-1.5B, is reported at 25.1 on the same LiveCodeBench window, up from 16.9 for its base.

DeepCoder-14B benchmarks: what "o3-mini level" means

Model	LiveCodeBench v5, Aug 2024 to Feb 2025	Codeforces rating	Codeforces percentile	HumanEval+
DeepCoder-14B-Preview	60.6	1936	95.3	92.6
DeepSeek-R1-Distill-Qwen-14B (its base)	53.0	1791	92.7	92.0
o3-mini 2025-01-31 (low)	60.9	1918	94.9	92.6
o1 2024-12-17 (low)	59.5	1991	96.1	90.8
DeepSeek-R1	62.8	1948	95.4	92.6

Sources: DeepCoder-14B-Preview model card and Together AI. All rows are as reported by DeepCoder's authors.

"o3-mini level" means o3-mini on low. DeepCoder is 0.3 points below that setting on LiveCodeBench, and 2.2 below DeepSeek-R1 in the same table.
The Codeforces ratings only compare inside this table. Agentica estimated them on Qwen's CodeElo set, 408 problems from 57 contests. DeepSeek's own card rates the same base model at 1481, against 1791 here: same model, different harness, 310 points apart.
The 60.6 needs 64K tokens of room. Capped at 16K, DeepCoder scores 45.6 against its base model's 50.2, because its long answers get cut off. It pulls ahead at 32K, 57.9 against 53.0, and reaches 60.6 at 64K.

Is there a newer DeepCoder?

Not as of April 8, 2026, a year to the day after the first one. Agentica's Hugging Face organisation lists five models: DeepScaleR-1.5B-Preview (January 2025), the two DeepCoders (April 2025), and DeepSWE-Preview with a verifier (July 2025). The rLLM project's posts since then announce no new DeepCoder. Both DeepCoder repositories were last changed in April and May 2025.

DeepSWE-Preview is Agentica's 32B software-engineering agent, built on Qwen3-32B: a different model for a different job, and not a DeepCoder. Naming things is still the hard part.

Is DeepCoder-14B still worth running in 2026?

People still look for it, mostly to run it locally. On Google Trends, searches for "deepcoder" jumped when it came out a year ago and have held at a lower, steady level since, and the related searches it marks as breakout include "deepcoder 14b" and "ollama deepcoder".

Searches are not a quality measure, and the published scores cannot settle it either. LiveCodeBench is versioned by problem date, and DeepCoder's 60.6 covers August 2024 to February 2025. Newer model cards tend to report LiveCodeBench v6, added in April 2025 with later problems, so I am not lining such numbers up. DeepCoder's test problems have all been public since February 2025, so a newer model may have trained on them. That is my inference.

DeepCoder-14B was trained and measured only on standalone programming problems with tests, not on repository edits or tool calls. For those, Agentica built a different, larger model. My judgement: it is worth running if you want a small MIT-licensed model for algorithmic work, offline, and you will compare it with a newer small model on your own tasks. I would not make it the engine of a coding agent.

How to run DeepCoder-14B: Hugging Face and Ollama

On Hugging Face the weights are 32-bit floats in twelve shards, about 59 GB by my arithmetic from the file sizes. The card lists vLLM, Text Generation Inference, SGLang and TensorRT-LLM for serving, and recommends no system prompt, temperature 0.6, top_p 0.95 and max_tokens of at least 64,000.

On Ollama, ollama run deepcoder pulls the 14B build, Q4_K_M at 9.0 GB, with the card's temperature and top_p preset. There are also q8_0 (16 GB), fp16 (30 GB) and deepcoder:1.5b (1.1 GB) builds. The catch is context. Ollama lists a 128K window for the model, but on most machines the server's default context is far shorter than the 64K the card asks for. Set OLLAMA_CONTEXT_LENGTH=64000 before ollama serve, and confirm the context the model actually got before judging its answers.

By my arithmetic from the model's configuration (48 layers, 8 KV heads of 128 dimensions), a 16-bit KV cache takes 196,608 bytes per token, about 12.6 GB for 64,000 tokens, on top of 9.0 GB of weights. That makes a 24 GB GPU about the floor for the full setting. I have not run it.

Prove an older model on your own tasks first

I have made this decision on cost rather than age. On the agent workflows behind my Auto-RAG and MerlinX work, the cheaper orchestrator model was first shown on a frozen evaluation suite, at 95% confidence, to be not worse than the model it replaced. The rollout then cut cost per conversation by 73% across 840 agents with no measurable quality loss.

That is the standard I would hold DeepCoder-14B to. A vendor table from April 2025 describes problems up to February 2025. A frozen set of your own tasks describes your work now.

What I would test before using DeepCoder-14B

The card's settings first. Reproduce 64K, temperature 0.6 and top_p 0.95, then the context your hardware allows, counting answers that stop before the code.
Problems no model has seen. Private ones you write are best. Public 2025 problems may be in a newer model's training data.
Time per solved problem. Long reasoning is slow on one GPU, and no benchmark table shows it.
The licence chain. DeepCoder and its base are MIT, and DeepSeek's card notes that the base derives from Qwen2.5, originally Apache 2.0.

Both DeepCoders arrived with a headline larger than the evidence behind it. The difference is that you can download the second one and check.

Sources

Written by Eslam Hamed, a Technical Lead and AI Solutions Architect working on enterprise RAG, agentic AI, and model selection.