Best Open-Source LLMs for Coding in 2026: Gemma 4 26B A4B IT and Gemma 4 31B IT

Quick Answer

Google’s Gemma 4 26B A4B IT is the top pick as of September 2026: with no public benchmark yet, the ordering is newest release first, then downloads, and it wins the downloads tiebreak among the newest releases [1][2][3][5][7]. In order, the ranking is Gemma 4 26B A4B IT (Google), Gemma 4 31B IT (Google), NVIDIA Nemotron 3 Nano 4B BF16 (NVIDIA), Qwen3.5-0.8B (Qwen), and Qwen3.5-9B (Qwen) [1][2].

Key Takeaways

• Ranking note: newest release first, then downloads; all candidates have “no public benchmark yet.” Scope admits only general chat LLMs with chat templates from labs with a published paper or leaderboard record, or other publishers above 10,000 downloads in the last 30 days, under text-generation or image-text-to-text tags; any-to-any generators, vision-first models, OCR, grounding, driving and other specialists are excluded. [1][2][3][5][7]
• Google’s Gemma 4 26B A4B IT ranks first because its March 11, 2026 release ties Google’s Gemma 4 31B IT for recency, while its downloads break the tie—not because of demonstrated coding superiority. [3][7] Its 25.8B parameters require 12.9 GB for quantized weights alone, estimated (params x 0.5 bytes); its license is Apache-2.0. [3]
• Google’s Gemma 4 31B IT ranks second and ships 62.5 GB of BF16 weights under Apache-2.0. [7] Its 31.3B parameters imply 15.65 GB for quantized weights alone, estimated (params x 0.5 bytes). [7]
• NVIDIA’s NVIDIA Nemotron 3 Nano 4B BF16 ranks third, with a March 7, 2026 release and 7.9 GB of BF16 weights. [5] Its 4B parameters imply 2 GB for quantized weights alone, estimated (params x 0.5 bytes); its license is nvidia-nemotron-open-model-license. [5]
• Qwen’s Qwen3.5-0.8B ranks fourth, with a February 28, 2026 release, 873M parameters and 1.7 GB of BF16 weights. [1] Apache-2.0 licensing and the modest weight footprint make it a candidate to evaluate when local memory is constrained; coding quality remains unranked by public benchmarks. [1]
• Qwen’s Qwen3.5-9B ranks fifth because its February 27, 2026 release predates the other candidates. [1][2][3][5][7] Its 9.7B parameters imply 4.85 GB for quantized weights alone, estimated (params x 0.5 bytes), versus 19.3 GB of BF16 weights; its license is Apache-2.0. [2]

How do these local coding LLMs compare on specifications and benchmark availability?

Which open-weight LLMs should you consider for local coding?

1. Gemma 4 26B A4B IT

Gemma 4 26B A4B IT by Google ranks first because it shares the newest release date among the candidates and wins the download tiebreak.[3][7] Published on March 11, 2026, it recorded 11,073,424 downloads over the reporting window, compared with 9,263,106 for Google’s Gemma 4 31B IT.[3][7] The ordering basis is newest release first, then downloads. Coding status: no public benchmark yet; the position does not establish superior coding performance.

The model has 25.8 billion parameters, a context length of 256K tokens, and BF16 weights totaling 51.6 GB.[3] Google distributes it under the Apache-2.0 license.[3] For local hardware planning, the four-bit weight payload is estimated (params x 0.5 bytes) at 12.9 GB, using the cited parameter count.[3] Treat that figure as a weight-only estimate, not a complete GPU memory or system RAM requirement. A specific hardware fit remains unverified.

Use Gemma as a candidate for local coding evaluation: try repository questions, patch generation, and test writing against code you can check. Treat the advertised context length as a capacity specification rather than proof of reliable repository-wide reasoning. The practical caveat is the missing public coding benchmark: choose it after evaluating correctness on your workload, rather than interpreting download counts as coding results.

2. Gemma 4 31B IT

Gemma 4 31B IT by Google ranks second under the ordering rule “newest release first, then downloads.”[3][7] Both ranked Gemma models were first published on Hugging Face on March 11, 2026.[3][7] Gemma 4 31B IT recorded 9,263,106 downloads over the preceding 30 days as of September 25, 2026, versus its sibling’s 11,073,424.[3][7] Downloads resolve that release-date tie; they do not establish coding quality.

The model has 31.3 billion parameters, a 256K-token context window and 62.5 GB of BF16 weights, under the Apache-2.0 license.[7] For local hardware planning, quantized weight storage would be approximately 15.65 GB, estimated (params x 0.5 bytes), using the cited 31.3-billion-parameter count.[7] Treat that figure as a weights-only estimate when budgeting system RAM or GPU VRAM. A complete hardware requirement cannot be established from weight storage alone.

Consider the model for local coding-assistant evaluation where a permissive license and the advertised context capacity match your requirements. Evaluate it against your own repository tasks before committing hardware: check whether proposed patches build, pass tests and satisfy the requested behavior. The central caveat is no public benchmark yet in the stipulated comparison, so its ranking does not demonstrate superior coding accuracy or speed.

3. NVIDIA Nemotron 3 Nano 4B BF16

NVIDIA Nemotron 3 Nano 4B BF16 by NVIDIA ranks third under the ordering rule: newest release first, then downloads [3][5][7]. Its March 7, 2026 publication date [5] puts it behind the March 11 releases [3][7] and ahead of the February releases [1][2]. For coding, the qualification is straightforward: no public benchmark yet. Its position reflects release timing, so the ranking does not establish superior coding performance.

The model has 4B parameters, a 256K-token context window and 7.9 GB of BF16 weights [5]. From that parameter count, a quantized copy would require approximately 2 GB for 4-bit weights alone—estimated (params x 0.5 bytes) [5]. Treat that calculation as a starting point for local hardware planning. Neither the weight size nor the advertised context window establishes a complete GPU or system RAM requirement; a specific hardware fit remains unverified.

Consider the model for local coding evaluation when its weight footprint matches your deployment budget. Test representative repository tasks before adopting it as a coding assistant, and measure memory use at the context lengths you actually need. The model uses the NVIDIA Nemotron Open Model License (nvidia-nemotron-open-model-license) [5]; review those terms against your intended deployment before distributing weights or integrating it into a product.

4. Qwen3.5-0.8B

Qwen3.5-0.8B by Qwen ranks fourth under the ordering rule: newest release first, then downloads.[1][2][3][5][7] Its first publication on Hugging Face was February 28, 2026, placing it behind the Google and NVIDIA candidates and ahead of Qwen3.5-9B by Qwen, published February 27, 2026.[1][2][3][5][7] With no public benchmark yet, that position does not establish relative coding ability. Treat it as a candidate for local evaluation, with your own coding tasks deciding whether to keep it.

The model has 873 million parameters, a 256K-token context length, and 1.7 GB of BF16 weights, under the Apache-2.0 license.[1] For local hardware planning, quantized weight storage is approximately 0.44 GB at four-bit precision—estimated (params x 0.5 bytes), using the cited 873 million parameters.[1] Use that estimate as a weights-only budget when considering available GPU memory or system RAM; neither weight figure establishes the total memory needed to run the model.

A practical role is a compact local coding assistant to evaluate on short explanations, small edits, and routine snippets. Those are suggested trial tasks, with no measured success rate implied. The caveat is unverified coding quality: the advertised context length should not be treated as evidence of reliable repository-wide reasoning. Check generated changes against expected behavior before incorporating them into a project.

5. Qwen3.5-9B

Qwen3.5-9B by Qwen ranks fifth because its Hugging Face publication date, February 27, 2026, precedes the other ranked releases.[2][1][3][5][7] The ordering is newest release first, then downloads. Qwen3.5-9B has no public benchmark yet, so its position does not establish weaker coding performance. Its 9,781,405 downloads over the preceding 30 days indicate adoption, but cannot establish coding quality.[2]

The model has 9.7 billion parameters, a 256K-token context window, and 19.3 GB of BF16 weights under the Apache-2.0 license.[2] For local hardware planning, 4-bit weights would occupy approximately 4.85 GB, estimated (params x 0.5 bytes) from the parameter count.[2] Treat that figure as a weights-only budget. Leave additional RAM or VRAM available for the runtime and context rather than selecting hardware whose capacity merely matches the weight estimate.

Consider it for evaluating a local coding assistant against your own repository: request a patch, run the relevant tests, and review the diff. Start with focused functions and explicit acceptance criteria before expanding to broader changes. The caveat is the lack of a public coding benchmark: evaluate correctness on representative tasks before relying on generated changes.

How are models selected and ranked when there is no public benchmark yet?

When no public benchmark scores the candidates, we rank by newest release first, then downloads, and label each model “no public benchmark yet.” [1][2][3][5][7] The order reflects release dates and a popularity tiebreak, not demonstrated coding performance.

Eligibility is limited to general chat LLMs that ship a chat template, use Hugging Face’s text-generation or image-text-to-text pipeline tags, and come from labs with a published paper or leaderboard record, or other publishers exceeding 10,000 downloads in the last 30 days; any-to-any generators, vision-first models, OCR, grounding, driving and other specialists are excluded. [1][2][3][4][5][6][7]

The ranked candidates pass the release window of the last 12 months. [1][2][3][5][7] Google’s google/gemma-4-26B-A4B-it ranks first because its publication date, 2026-03-11, ties Google’s google/gemma-4-31B-it for the newest release, while its 11,073,424 downloads exceed the latter’s 9,263,106 over the last 30 days as of 2026-09-25. [3][7]

A public leaderboard can determine benchmark order only among models it evaluates together. An older scored model cannot outrank a newer unscored model on that basis. Model-card benchmark results, when included, are explicitly marked “self-reported.” Download counts break release-date ties; parameter counts never determine the order.

For a local coding deployment, read the ranking alongside weight storage, context length and license terms. Hardware-fit calculations must be labeled estimates. Neither popularity nor recency establishes how reliably a model will edit your repository or complete your coding tasks.

How much memory should you budget for local inference?

Budget for weights plus runtime headroom: the listed BF16 weights range from 1.7 GB [1] to 62.5 GB [7], but those figures are not complete inference memory requirements.

Qwen’s Qwen3.5-0.8B has 1.7 GB of BF16 weights and 873M parameters [1]; its quantized weight budget is approximately 0.44 GB at 4-bit, estimated (params x 0.5 bytes) [1]. Qwen’s Qwen3.5-9B has 19.3 GB of BF16 weights and 9.7B parameters [2]; allow approximately 4.85 GB at 4-bit, estimated (params x 0.5 bytes) [2].

Google’s gemma-4-26B-A4B-it has 51.6 GB of BF16 weights and 25.8B parameters [3]; allow approximately 12.9 GB at 4-bit, estimated (params x 0.5 bytes) [3]. Google’s gemma-4-31B-it has 62.5 GB of BF16 weights and 31.3B parameters [7]; allow approximately 15.65 GB at 4-bit, estimated (params x 0.5 bytes) [7].

NVIDIA’s NVIDIA-Nemotron-3-Nano-4B-BF16 has 7.9 GB of BF16 weights and 4B parameters [5]; allow approximately 2 GB at 4-bit, estimated (params x 0.5 bytes) [5].

Use the weight estimates as a starting point for hardware planning. Leave additional memory for the inference runtime, working buffers, and KV cache rather than matching GPU memory or system RAM exactly to the weight budget. Validate the intended quantization, prompt length, and concurrency in your chosen runtime before committing to hardware.

Which licenses apply to these models?

Apache-2.0 applies to the listed Qwen and Google models, while the NVIDIA model uses the NVIDIA Nemotron Open Model License (nvidia-nemotron-open-model-license). [1][2][3][5][7] Check the license for the exact model you plan to deploy before approving its use.

Qwen’s Qwen3.5-0.8B and Qwen3.5-9B both list Apache-2.0. [1][2] Google’s Gemma 4 26B A4B instruction-tuned model (google/gemma-4-26B-A4B-it) and Gemma 4 31B instruction-tuned model (google/gemma-4-31B-it) also list Apache-2.0. [3][7] For an engineering inventory, record each repository separately even when the license identifier matches. Keep the applicable license text with your deployment documentation, and review its requirements before modifying or redistributing weights.

NVIDIA’s NVIDIA-Nemotron-3-Nano-4B-BF16 lists nvidia-nemotron-open-model-license, so its license review needs to address that agreement. [5] Do not assume that an approval covering the Apache-licensed candidates also covers the NVIDIA candidate. Check the agreement against your intended use, including internal deployment, customer-facing services, modification and redistribution.

For deployment decisions, treat “open-source” in the article title as a starting point for evaluation. Base approval on the applicable license text and your intended use; record any conditions your team must follow alongside the selected model.

Frequently Asked Questions

Which model takes the first position for local coding?

Google’s google/gemma-4-26B-A4B-it takes the first position because its March 11, 2026 release ties Google’s google/gemma-4-31B-it, while its monthly downloads are higher: 11,073,424 versus 9,263,106. [3][7] The recommendation follows release date and a popularity tiebreak; coding performance remains unproven here, with no public benchmark yet. Its Apache-2.0 license is also relevant when evaluating it for a development workflow. [3]

How does the ranking work, and which models qualify?

Ranking note: newest release first, then downloads; download counts only break release-date ties. Scope: this ranking admits only general chat LLMs shipping a chat template, from labs with a published paper or leaderboard record or other publishers above 10,000 downloads in the last 30 days, using text-generation or image-text-to-text tags; any-to-any generators, vision-first models, OCR, grounding, driving and other specialists are excluded. [1][2][3][5][7] Every candidate has no public benchmark yet.

What is the full model order?

The order is Google’s google/gemma-4-26B-A4B-it, Google’s google/gemma-4-31B-it, NVIDIA’s nvidia/NVIDIA-Nemotron-3-Nano-4B-BF16, Qwen’s Qwen3.5-0.8B (Qwen), then Qwen’s Qwen3.5-9B (Qwen). [3][7][5][1][2] NVIDIA’s entry was published on March 7, 2026; the Qwen entries followed the opposite size order, appearing on February 28 and February 27, respectively. [5][1][2] Treat the sequence as an evaluation shortlist: no public benchmark yet establishes their relative coding ability.

How much memory should I budget for local weights?

For Google’s google/gemma-4-26B-A4B-it, 25.8B parameters imply 12.9 GB for four-bit weights, estimated (params x 0.5 bytes). [3] NVIDIA’s nvidia/NVIDIA-Nemotron-3-Nano-4B-BF16 has 4B parameters, implying 2 GB, estimated (params x 0.5 bytes). [5] Qwen’s Qwen3.5-9B (Qwen) has 9.7B parameters, implying 4.85 GB, estimated (params x 0.5 bytes). [2] Those calculations describe weight storage only; they do not establish a complete RAM requirement or guarantee that a particular GPU can run the model.

Which model should I evaluate on a machine with limited memory?

Start by evaluating Qwen’s Qwen3.5-0.8B (Qwen): its listed BF16 weights occupy 1.7 GB, compared with 19.3 GB for Qwen3.5-9B (Qwen). [1][2] Both list a context length of 256K tokens. [1][2] Use representative coding tasks to decide whether the smaller model meets your needs. No public benchmark yet establishes the coding tradeoff between these candidates, and the listed weight sizes do not establish complete hardware requirements.

Are all the models available under the same open-source license?

The licenses differ. Google’s Gemma candidates and Qwen’s candidates list Apache-2.0. [3][7][1][2] NVIDIA’s nvidia/NVIDIA-Nemotron-3-Nano-4B-BF16 lists the nvidia-nemotron-open-model-license. [5] Describe the shortlist as open-weight, and retain each model’s specific license name when documenting your choice. For a deployment decision, review the applicable license against your intended use rather than treating every downloadable model as having identical terms.

Sources

1. Qwen/Qwen3.5-0.8B model card (Hugging Face) — 2026-09-25
2. Qwen/Qwen3.5-9B model card (Hugging Face) — 2026-09-25
3. google/gemma-4-26B-A4B-it model card (Hugging Face) — 2026-09-25
4. Gemma 4 Technical Report — 2026-07-02
5. nvidia/NVIDIA-Nemotron-3-Nano-4B-BF16 model card (Hugging Face) — 2026-09-25
6. Nemotron Elastic: Towards Efficient Many-in-One Reasoning LLMs — 2025-11-20
7. google/gemma-4-31B-it model card (Hugging Face) — 2026-09-25
8. meta-llama/Llama-3.2-11B-Vision-Instruct model card (Hugging Face) — 2024-09-18
9. moonshotai/Kimi-K2.6 model card (Hugging Face) — 2026-04-14